WPS7799


  Policy Research Working Paper                  7799




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

                                Quy-Toan Do
                                 Shareen Joshi
                                Samuel Stolper




  Development Research Group
  Poverty and Inequality Team
  August 2016
Policy Research Working Paper 7799


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




  This paper is a product of the Poverty and Inequality Team, Development Research Group. It is part of a larger effort by
  the World Bank to provide open access to its research and make a contribution to development policy discussions around
  the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be
  contacted at Abonfield@worldbank.org.




         The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
         issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
         names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
         of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
         its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.


                                                       Produced by the Research Support Team
         Can Environmental Policy Reduce Infant Mortality?
                                                                                                    ∗
                  Evidence from the Ganga Pollution Cases

          Quy-Toan Do                      Shareen Joshi                          Samuel Stolper
           World Bank                Georgetown University                  University of Michigan




Keywords: environmental policy, development, pollution, infant mortality.

JEL Codes: Q53, Q56, C36
   ∗ We are grateful to Prashant Bharadwaj, Jishnu Das, Ram Fishman, Garance Genicot, Mark Giordano, Susan

Godlonton, Rema Hanna, Hanan Jacoby, Remi Jedwab, Guido Kuersteiner, Shomik Lall, Arik Levinson, Rohini Pande,
Martin Rama, Martin Ravallion, John Rust, Simone Schaner, George Shambaugh, Shinsuke Tanaka, and Jennifer Tobin
for useful comments. Many thanks to Mr. Jairam Ramesh and Mr. M.C. Mehta for helpful discussions. Thank you as
well to Michael Greenstone and Rema Hanna for sharing their pollution data, to Sam Asher and Paul Novosad for sharing
their weather data, and to Shreya Shah for valuable qualitative research. The findings, interpretations, and conclusions
expressed in this work do not necessarily reflect the views of the World Bank, its Board of Executive Directors, or the
governments they represent.
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,

and Brainerd and Menon 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

(Hanna and Oliva 2015) and high marginal costs of pollution abatement (Davis 2008), they are also

likely driven by rent-seeking behavior (Duflo et al. 2013) and market failures (Jalan and Somanathan

2008), hence defining the scope of government regulation.

   India provides a compelling setting in which to study developing-country environmental health

and policy. The World Health Organization estimates that over three in every 1,000 Indian children

under five years old in 2004 died because of water pollution (WHO 2004). India’s Central Pollution

Control Board reports that sewage treatment capacity amounts to only 21 percent of the estimated

daily sewage load in Indian cities (CPCB 2017a), and the gap between load and treatment capacity

is expanding (Daily Mail 2015). At the same time, the flagship policy for addressing water pollution

in India – the National River Conservation Plan (NRCP), established in 1986 and still active today –

has failed to improve water quality (Greenstone and Hanna 2014).

   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 rulings – which mandated pollution cleanup by the tanning industry concentrated

along the Ganga River in Kanpur, Uttar Pradesh – marked 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


                                                   2
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 economically-

meaningful 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



                                                    3
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 afore-

mentioned 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 overidentifi-

cation tests fail to reject the null hypothesis that the rulings are a valid instrument for river pollution –

i.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.

Section 6 concludes.



1     The Ganga Pollution Cases

In the aftermath of several decades of population and industrial growth, India’s rivers are heavily

polluted – particularly 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
   1 India’s official water quality criteria have five tiers (A-E), each of which includes an acceptable range for multiple

water quality measures, including BOD, dissolved oxygen, total coliforms, and pH. Outdoor bathing is tier B; tier A
corresponds to untreated (but disinfected) drinking water.


                                                           4
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



                                                    5
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).



2    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 (1986-1995). 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.




                                                    6
2.1    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 short-

term 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

an infant’s life (United Nations Children’s Fund 2015). Globally, neonatal mortality accounts for over

half of all infant deaths (World Development Indicators 2017); in our own data from India, its share of

under-one mortality is 70 percent. Furthermore, several existing studies of pollution – in both water

and air – and child health show that the largest impacts occur within one month of birth (Chay and

Greenstone 2003; Gluckman et al. 2008; Currie and Almond 2011; and Brainerd and Menon 2014).

While neonatal mortality is not a complete measure of the health costs imposed by water pollution, it

nonetheless represents a very large loss of life, especially in the Indian context.

   Our infant health data come from the Reproductive and Child Health II (RCH-2) module of the

District-Level Household Survey II (DLHS-2), a national demographic survey conducted in two phases

from 2002 to 2005. In the RCH-2 module, mothers report age and survival for all of their children;

from these birth histories, we create a panel of child-month mortality in the first year. The full

national dataset spans the years 1967-2004. The first four years of this range includes 2, 0, 0, and

6 reported births nationally; starting in 1971, the reported birth count rises roughly monotonically

from 42 towards its peak of 80,961 in 1998. Because of the lack of data from 1967-1970, we drop these

years from the sample. This leaves us with with 1,393,330 birth observations from the years 1971-2004,

647,865 of which are to mothers living within the Ganga Basin (240 districts) at the time of survey.


2.2    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


                                                    7
Hanna (2014), from a combination of CPCB online and print records. The dataset spans the whole of

India and runs from 1986-2004; an observation is a pollution monitor-month.

    The national dataset contains information from 470 monitors situated along 162 rivers in 28 states.

Excluding all areas outside the Ganga Basin leaves us with 101 unique pollution monitors along 29

rivers, in 62 districts representing 10 states; these are mapped in Figure 2. Over our 19-year period

of observation, as many as 46 different measures of water quality are recorded at these monitoring

stations, but only a few measures are consistently recorded over the entire sample time frame. The

data are noisy, suggesting significant measurement error, and the panel is imbalanced. To mitigate these

problems, we collapse monitor-level observations to the district level and construct moving averages

of the data over a four-month window. These steps produce 8,725 district-month observations in the

Ganga Basin with at least one non-missing value from our pollution measures of interest.

    Our main indicator of river quality is BOD. This common, broad-based indicator of water pollution

captures the amount of dissolved oxygen needed by water-borne, aerobic organisms to break down

organic material present at a certain temperature (usually 20 degrees Celsius) and over a specific time

period (usually five days). Its units are milligrams of oxygen consumed per liter (mg/l). Reduction of

BOD is the primary goal of waste treatment plants in general (Brown and Caldwell 2001), but BOD is a

particularly good choice for pollution measurement in the setting of Kanpur. Pollution from the tanning

process primarily comes from two sources: the animal hides themselves, and the chemicals used to tan

them. Both of these sources contain large amounts of organic matter, resulting in abnormally high

BOD levels in tannery effluent. According to the United Nations Industrial Development Organization

(UNIDO), effluent discharge into surface water typically is required to have BOD below 30-40 mg/l,

while the typical BOD in raw tannery effluent is approximately 2,000 (UNIDO 2011).

    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 munici-

palities 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-
   2 In addition to this BOD standard, suitability for outdoor bathing requires a total coliforms concentration of less

than 5 per ml, pH between 6.5 and 8.5, and dissolved oxygen levels exceeding 5 mg/l.



                                                          8
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 sta-

tistically 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 and Greenstone 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


2.2.1    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.




                                                           9
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.


2.3     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



3     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 verdict – to 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
   5 We are unable to weight tributaries by their contribution to downstream water flow, due to a lack of data on

volumetric flow rates at specific locations.
   6 We do not have any other information about the placement, utilization or other operational details of this CETP,

or others like it, in India. All the results in this paper however, are robust to the exclusion of these policy controls.



                                                           10
on the mechanisms through which environmental regulation affects health.


3.1     Difference-in-differences

We specify the reduced-form impact of the Supreme Court verdict on neonatal mortality as follows:


                                    M ortalityidt = α1 + α2 Tdt + Xidt η + eidt                                         (1)


where M ortalityidt is a dummy variable indicating whether a child i, born in district d, in year-month

t, died within the first month of life. Tdt captures the incidence of policy – the Mehta vs. Union

of India court decisions in our case – and takes a value of one in affected — or “treated”— districts

after October 1987 and zero otherwise. Xidt 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:

                                     P ollutiondt = β1 + β2 Tdt + Xidt θ +         idt                                  (2)


where P ollutiondt 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 (Tdt ,                  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

prior precedent in India, and because M.C. Mehta’s choice of Kanpur was motivated primarily by

its central location and the salience of pollutants coming from its tanneries (M.C. Mehta, personal

communication, 12/16/2014). Thus, Kanpur was not a priori chosen on the basis of temporal trends
   7 There is no pollution monitor in Fatehpur district, so the treatment group contains only three districts in regressions

involving pollution data.




                                                            11
in pollution, health, or citizen involvement. In fact, there is no evidence of any local movement to

reduce pollution in the city in the mid-1980s (Jaiswal 2007).

   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.


3.2    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

consequences of contamination. Policy research in the environment-development literature frequently

does not attempt to disentangle the relative importance of the reduced-contamination channel from

that of other possible channels (e.g., Greenstone and Hanna 2014 and Watson 2006).

   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 urban-

ization (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


                                                  12
upstream counterpart, as defined in Section 2. The logic of this instrument leverages the unidirec-

tional 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:


                       M ortalityidt = γ1 + γ2 Tdt + γ3 P ollutionu
                                                                  dt + Xidt φ +   idt ,                 (3)


where P ollutionu
                dt is the pollution reading upstream of district d in year t. The primary parameter

of interest in equation (3) is γ3 , the impact of upstream pollution on downstream mortality. The

point estimate of γ3 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 (2002, 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 γ3 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,


                       P ollutiondt = γ1 + γ2 Tdt + γ3 P ollutionu
                                                                 dt + Xidt φ +    idt                   (4)


and, as a second stage,


                       M ortalityidt = γ1 + γ2 Tdt + γ3 P ollutiondt + Xidt φ +   idt .                 (5)


Coefficient γ3 in equation (5) now provides the estimated impact of river pollution on neonatal mortality

in the average district. This point estimate provides a reference point for the magnitude of pollution’s

burden on infant health. Combining this with the spillover mortality impact in the next district

downstream identified from equation (3), we have a summary estimate of the external infant mortality

cost imposed by river pollution.

   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



                                                    13
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 Tdt and P ollutionu
                                                                                             dt as

instruments on the one hand and the regression using only P ollutionu
                                                                    dt on the other hand.
                                                                                         8




4          Results

We report summary statistics for each analysis variable in Table 1, comparing the treatment group to

the control group. The top panel tabulates statistics from variables observed at the child level. The

bottom panel, meanwhile, describes those variables that we measure at the district level.

     We highlight several noteworthy observations from Table 1. District-level neonatal and infant mor-

tality rates over our observed time period are somewhat larger (8 percent and 11 percent, respectively)

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.


4.1        Difference-in-differences: a graphical representation

We begin our analysis by graphically inspecting our difference-in-differences (DD) assumption of par-

allel 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.



                                                         14
control group. Pre-ruling summary statistics for treatment and control are provided in Appendix Table

A1.

    Figure 3 contains two panels; the top panel shows neonatal mortality averages from the full period,

while the bottom panel uses data within a smaller, symmetric 15-year window (1980-1994) surrounding

the Supreme Court verdicts. In both panels, the vertical distance between the treatment and control

group means is smaller after the verdicts than before them. The smaller scale of Panel B allows for a

closer look, which shows parallel decreasing pre-trends, a roughly-constant slope in the control group

over the full period, and a drop in treatment-group mortality in the aftermath of the verdicts.

    Figure 4 provides analogous results for river pollution, using a dummy for exceedance of the bathing

quality standard (BOD>3) as the dependent variable. Since this specification requires pollution data,

the full time period available to us is 1986-2004. The top panel utilizes the full period, while the

bottom panel restricts the estimation sample to 1986-1994 – a “short-run” post-period. Here, the

outcome variable is somewhat noisy in the treatment group. In the short pre-period available to us

for pollution, however, treatment- and control-group trends are parallel. Meanwhile, the likelihood

of BOD exceeding the 3-mg/l cutoff drops significantly for the treatment group in the post-period,

indicating a large DD estimate of the impact of the ruling on water quality.

    While in principle it is possible to compare our treatment group to a single nearby (but not

downstream) control district, inherent noise in our infant mortality and pollution data makes such a

comparison imprecise. We illustrate this in Appendix Figure 1 by plotting average mortality (Panel

A) and pollution (Panel B) in the treatment group and in a single district (Kannauj) just upstream
                        10
of Kanpur Nagar.             These graphs document a large time-series variance in outcomes of interest.

Using the entirety of the Ganga Basin in our control group allows us to smooth our counterfactual

significantly.11


4.2     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
  10 The district of Kannauj is immediately upstream of Kanpur Nagar. The two Kannauj pollution monitors in our

data lay approximately 93 kilometers upstream of the tanneries.
  11 We also test the robustness of the graphical DD strategy to the definition of the treatment group, by graphing

treatment- and control-group average outcomes after redefining the treatment group to only include Kanpur Nagar
district itself. The results are depicted in Appendix Figure A2; they do not change in any qualitatively meaningful way.




                                                          15
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 separate out these one-year mortality results into impacts

in the first month of life vs. in months two through twelve. Given a baseline neonatal mortality

risk of 12.1 percent in the treatment group prior to the verdicts, columns 2 and 5 indicate large

impacts, with statistically significant point estimates of -0.024 (or 19.5 percent) and -0.020 (or 16.2

percent), respectively. Meanwhile, the estimated mortality impact in months two through twelve is a

much smaller and statistically insignificant -0.001 in both column 3 and column 6. Together, Tables

1 and 2 reveal that most infant mortality events are neonatal and that the environmental policy in

question here is associated with large-magnitude reductions in neonatal mortality – on the order of

16-20 percentage points. The adverse effect of pollution beyond the first month is more likely to be

captured by morbidity, which we do not observe over time. We thus focus exclusively on neonatal

mortality for the remainder of this paper, touching on possible mechanisms of exposure in Section 5.

    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
  12 For example, column 1 indicates that, after the verdicts, the probability of infant death in a given month drops by

0.005 in the Kanpur region. This translates to a 28.9 percent reduction in the aggregate risk of infant mortality. The
significance of point estimates in Table 2 is unaffected by choice of weights.




                                                          16
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 in significant 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.

   In column 7, we test whether neonatal mortality impacts of regulation persist downstream of our

defined treatment group. We define three downstream groups. ‘Near Downstream’ denotes the near-

est group downstream and includes Pratapgarh, Kaushambi, and Allahabad districts. ‘Intermediate

Downstream’ is the next nearest group and consists of Sant Ravidas Nagar, Mirzapur, and Varanasi

districts. ‘Far Downstream’ collects the remaining 11 Indian districts through which the Ganga River

travels beyond Varanasi. We estimate coefficients on the interaction between dummies for each of these

groups and the post-ruling dummy. Only the ‘Near Downstream’ coefficient is statistically significant.

In fact, it is significantly larger than the main treatment-group coefficient (-0.32 vs. -0.20), which

may be explained by the location of Kanpur’s tanneries very near to the downstream district border.

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.


4.3    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 reports regression results with our primary pollutant, BOD,

while Panel B shows results for other relevant pollutants.

   The BOD results displayed in columns 1-6 uniformly imply a significant drop in pollution. Our

baseline estimate, shown in column 1 and corresponding to the full time-period of available pollution

data, implies a 53.4 percentage-point drop in the likelihood of exceeding the bathing class BOD stan-

dard after the Supreme Court rulings. Prior to the ruling, the likelihood of exceedance in treated

districts was 100 percent. The drop in exceedance is consistently observed throughout the post-verdict

period, as shown in columns 2-4. It is also robust to an expansion of the control group to include all

of India (column 5). Meanwhile, downstream impacts in column 6 mirror those from Table 3, column

7: the nearest group of districts downstream of the treatment group is associated with a large and


                                                    17
significant average drop in exceedance of the water quality standard, and the point estimate monoton-

ically 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. Figure 5 provides kernel densities of the residuals from regressing log- BOD

on weather controls and district and year-month fixed effects. Panel A suggests that the verdicts are

associated with significant, yet non-uniform, drops in BOD levels; the central peak shifts downwards,

but the left tail is relatively unchanged. In contrast, Panel B illustrates a symmetric tightening of the

log-BOD distribution around the mean. Put together, the two kernel density plots paint a picture of

treatment-group reductions in BOD that are relatively concentrated in the middle to upper half of the

distribution.

    Panel B of Table 4 provides estimates of the impact of other pollutants that are relevant to the

tanning industry and the Supreme Court verdicts. Calcium, sulfate, and chloride ions are all by-

products of the tanning process; all drop significantly in the aftermath of the rulings (columns 1, 2,

and 3). FCOLI, meanwhile, is a strong indicator of domestic pollution and thus should be observed

to drop if the second of the two Supreme Court verdicts was successful. We find no evidence of such

a drop (column 4).


4.4     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
  13 The 75-km lower bound minimizes the risk of off-river spatial correlation, while the 200-km upper bound excludes

locations that are so far upstream as to not physically affect downstream pollution levels. We test the robustness of the


                                                          18
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 non-

missing 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.




                                                           19
    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 We further test for the role of pollution in policy

impacts by re-estimating the 2SLS model in columns 5-8 with the same parameters as in columns 1-4,

respectively, except with two instruments instead of one. A rejection of the null (i.e., that the policy

instrument is valid) is evidence of either (a) multiple channels of policy impact or (b) different local

average treatment effects being captured by the two instruments. A failure to reject, on the other

hand, does not formally rule out the existence of other mechanisms, but it does support the notion

that reductions in pollution were central to the observed declines in mortality.

    The results of the over-identification tests are tabulated as p-values for the associated C -statistic

in columns 5-8. These p-values are statistical measures of the difference between the dose-response

coefficients from 1-IV and 2-IV specifications. The naked-eye version of the test is therefore a pairwise

comparison of coefficients on 1[BOD > 3] between columns 1 and 5, 2 and 6, 3 and 7, and 4 and 8.

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.



5     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 (1986-2004), 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.



                                                            20
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 pollu-

tion 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 contami-

nation 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.

Air quality in Delhi, India, for example, averaged 128 µg/m3 of fine particulate matter (PM2.5) in 2015

– ten times more than in Washington D.C. (Washington Post 2016) and significantly worse than in

2010. In that same five-year period, the number of classified ’polluted river stretches’ in India doubled

from 150 to 302, and the gap between sewage load and sewage treatment capacity expanded (Daily

Mail 2015). Why, if pollution is so costly, does it persist at high levels?

    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,



                                                           21
    High marginal costs of abatement and political economic distortions are two other possible expla-

nations 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 (2002,

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 environ-

mental 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). Further understanding of what made the Ganga

Pollution Cases successful might therefore shed light on what types of institutional arrangements lead

to successful policy implementation.



6     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.


                                                         22
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.




                                                    23
References

 [1] Alley, Kelly D. (2002). On the Banks of the Ganga: When Wastewater Meets a Sacred River.

    University of Michigan Press.

 [2] Angelucci, Manuela and Orazio Attanasio (2013). “The Demand for Food of Poor Urban Mexi-

    can Households: Understanding Policy Impacts Using Structural Models.” American Economic

    Journal: Economic Policy 5(1): 146-178.

 [3] APHA (1992). Standard Methods for the Examination of Water and Wastewater. 18th ed. Amer-

    ican Public Health Association, Washington, DC.

 [4] Beg, K. R. and S. Ali (2008). “Chemical contaminants and toxicity of Ganga River sediment

    from up and down stream area at Kanpur.” American Journal of Environmental Sciences 4(4):

    362-366.

 [5] Brainerd, Elizabeth and Nidhiya Menon (2014). “Seasonal Effects of Water Quality: The Hidden

    Costs of the Green Revolution to Infant and Child Health in India.” Journal of Development

    Economics 107: 49-64.

 [6] Brown and Caldwell (2001). “A Guidebook for Local Governments Developing Regional Water-

    shed Protection Plans: Appendix B. Description of Commonly-Considered Water Quality Con-

    stituents.” Prepared for the Georgia Environmental Protection Division.

 [7] Central Pollution Control Board (2009). “Ganga Water Quality Trend”. Ministry of Environment

    and Forests, Government of India.

 [8] Central Pollution Control Board (2017a). “Status of STPs”. Ministry of Environment and Forests,

    Government of India. <http://cpcb.nic.in/statusSTP.php>. Accessed on 7/8/2016.

 [9] Central Pollution Control Board (2017b). “Water Quality / Pollution”. Central Pol-

    lution Control Board,      Ministry of Environment and Forests,           Government of India.

    <http://cpcb.nic.in/water.php>. Accessed on 7/8/2016.

[10] Chay, Kenneth and Michael Greenstone (2003). “The Impact of Air Pollution on Infant Mortality:

    Evidence from Geographic Variation in Pollution Shocks Induced by a Recession.” Quarterly

    Journal of Economics 118(3): 1121-1167.

[11] Cheremisinoff, Nicholas P. (2001). Handbook of pollution prevention practices. CRC Press.

[12] Cifuentes, Enrique, Mariana Gomez, Ursula Blumenthal, Martha M. Tellez-Rojo, Isabel Romieu,

    Guillermo Ruiz-Palacios, and Silvia Ruiz-Velazco (2000). “Risk Factors for Giardia Intestinalis


                                                24
    Infection in Agricultural Villages Practicing Wastewater Irrigation in Mexico.” American Journal

    of Tropical Medicine and Hygiene 62(3): 388-392.

[13] Currie, Janet. and Douglas Almond (2011). “Human capital development before age five.” Hand-

    book of Labor Economics 4: 1315-1486.

[14] Daily Mail (2015). “They’re Raising a Stink: Polluted Rivers in the Country Have Doubled over

    Past Five Years.” April 3rd, 2015. Written by Akashi Vashishtha.

[15] Davis, Lucas (2008). “The Effect of Driving Restrictions on Air Quality in Mexico City.” Journal

    of Political Economy 116(1): 38-81.

[16] Duflo, Esther, Michael Greenstone, Rohini Pande, and Nicholas Ryan (2013). “Truth-Telling by

    Third-Party Auditors and the Response of Polluting Firms: Experimental Evidence from India.”

    Quarterly Journal of Economics 128(4): 1499-1545.

[17] Durai, G. and M. Rajasimman (2011). “Biological Treatment of Tannery Wastewater - A Review.”

    Journal of Environmental Science and Technology 4: 1-17.

[18] Ebenstein, Avraham (2012). “The Consequences of Industrialization: Evidence from Water Pollu-

    tion and Digestive Cancers in China.” Review of Economics and Statistics 94(1): 186-201.

[19] Eichenbaum, Martin S., Lars P. Hansen, and Kenneth J. Singleton (1988). “A Time Series Anal-

    ysis of Representative Agent Models of Consumption and Leisure Choices under Uncertainty.”

    Quarterly Journal of Economics 103(1): 51-78.

[20] Field, Erica, Rachel Glennerster, and Reshmaan Hussam (2011). “Throwing the Baby Out with

    the Drinking Water: Unintended Consequences of Arsenic Mitigation Efforts in Bangladesh.”

    Mimeo.

[21] Gamper-Rabindran, Shanti, Shakeeb Khan, and Christopher Timmins (2010). “The Impact of

    Piped Water Provision: A Quantile Panel Data Approach.” Journal of Development Economics

    92: 188-200.

[22] Garg, Teevrat, Stuart Hamilton, Jacob Hochard, Evan Plous, and John Talbot (2016). “Not So

    Gently Down the Stream: River Pollution and Health in Indonesia.” Grantham Research Institute

    on Climate Change and the Environment, Working Paper #234.

[23] Gluckman, Peter D., Mark A. Hanson, Cyrus Cooper, and Kent L. Thornburg (2008). “Effect of

    in utero and early-life conditions on adult health and disease.” New England Journal of Medicine

    359(1):61-73.


                                                25
[24] Graff-Zivin, Joshua, Matthew Neidell, and Wolfram Schlenker (2011). “Water Quality Violations

    and Avoidance Behavior: Evidence from Bottled Water Consumption.” American Economic Re-

    view: Papers & Proceedings 101(3): 448-453.

[25] Greenstone, Michael and Rema Hanna (2014). “Environmental Regulations, Air and Water Pol-

    lution, and Infant Mortality in India.” American Economic Review 104(10): 3038-3072.

[26] Greenstone, Michael and B. Kelsey Jack (2015). “Envirodevonomics: A Research Agenda for an

    Emerging Field.” Journal of Economic Literature 53(1): 5-42.

[27] Hanna, Rema and Paulina Oliva (2015). “Moving up the Energy Ladder: The Effect of an In-

    crease in Economic Well-Being on the Fuel Consumption Choices of the Poor in India.” American

    Economic Review Papers & Proceedings 105(5): 242-246.

[28] The Hindu (2016). “Curb on Tanneries During Magh Mela.” January 5th, 2016.

[29] Hanna, Rema and Paulina Oliva (2015). “The Effect of Pollution on Labor Supply: Evidence from

    a Natural Experiment in Mexico City.” Journal of Public Economics 122: 68-79.

[30] International Standards Organization (2003). “Water Quality–Determination of the Biochemical

    Oxygen Demand After N Days.” <http://www.iso.org/obp/ui/#iso:std:iso:5815:-1:ed-1:v1:en>.

[31] Jaiswal, Rakesh K. (2007). Ganga action plan – a critical analysis. Kanpur: Eco Friends. Mimeo.

    <http://www.ecofriends.org/main/eganga/images/Critical%20analysis%20of%20GAP.pdf>.

[32] Jalan, Jyotsna and E. Somanathan (2008). “The Importance of Being Informed: Experimental

    Evidence on Demand for Environmental Quality.” Journal of Development Economics 87: 14-28.

[33] Jayachandran, Seema (2008). “Air Quality and Early-Life Mortality: Evidence from Indonesia’s

    Wildfires.” Journal of Human Resources 44(4): 916-954.

[34] Kremer, Michael, Jessica Leino, Edward Miguel, and Alix Peterson Zwane (2011). “Spring Clean-

    ing: Rural Water Impacts, Valuation, and Property Rights Institutions.” Quarterly Journal of

    Economics 126(1): 145-205.

[35] Lipscomb, Molly and A. Moshfiq Mobarak (2017). “Decentralization and Water Pollution

    Spillovers: Evidence from the Redrawing of County Boundaries in Brazil.” Review of Economic

    Studies 84(1): 464-502.

[36] Madhya     Pradesh       Pollution   Control        Board   (2017).   “Ganga   Action    Plan”.

    <http://www.mppcb.nic.in/gap.htm>. Accessed on 7/8/2016.



                                                    26
[37] Mehta, M.C. (2009). In the Public Interest. Prakriti Publications: New Delhi.

[38] Ministry of Environment and Forests, Government of India. (2013). “Annual Report 2012-13”.

    New Delhi, India.

[39] Murty, M. N. and Surender Kumar (2011). “Water Pollution in India: An Economic Appraisal.”

    India Infrastructure Report 2011: Water Policy and Performance for Sustainable Development.

[40] Ravallion, Martin and Jyotsna Jalan (2003). “Does Piped Water Reduce Diarrhea for Children in

    Rural India.” Journal of Econometrics 112: 153-173.

[41] Sigman, Hilary (2002). “International Spillovers and Water Quality in Rivers: Do Countries Free

    Ride?” American Economic Review Papers and Proceedings 92(4): 1152-1159.

[42] Sigman, Hilary (2005). “Transboundary Spillovers and Decentralization of Environmental Poli-

    cies.” Journal of Environmental Economics and Management 50: 82-101.

[43] Singh,      Indra   B.   (2007).    “The    Ganga     River”.    In    Avijit   Gupta     (ed.),

    Large Rivers: Geomorphology and Management. John Wiley & Sons: West Sussex, UK.

[44] Singh, Amit (2014). “Judicial Activism on Environment in India”. Working Paper.

[45] Stout, Matthew D., Ronald A. Herbert, Grace E. Kissling, Bradley J. Collins, Gregory S. Travlos,

    Kristine L. Witt, Ronald L. Melnick, Kamal M. Abdo, David E. Malarkey, and Michelle J. Hooth

    (2009). “Hexavalent Chromium Is Carcinogenic to F344/N Rats and B6C3F1 Mice after Chronic

    Oral Exposure.” Environmental Health Perspectives 117(5):716-722.

[46] Suresh, Babu S.V., Bharat L. Seth, Ajit Chak, Vasudha K. Dixit, and Sunita Narain (2007).

    Sewage Canal: How to Clean the Yamuna. New Delhi: Centre for Science and Environment.

[47] Supreme Court of India (1985). Record of Proceedings In the Supreme Court of India, Civil

    Original Jurisdiction, Writ Petition (Civil) No. 3727/1985.

[48] United Nations Childrens’ Fund (2015). “Levels and Trends in Child Mortality: Report 2015.”

[49] United Nations Industrial Development Organization (2011). “Introduction to Treatment of Tan-

    nery Effluents: What Every Tanner Should Known about Effluent Treatment.”

[50] Washington Post (2016). “The Country with the Worst Air Pollution Is Not the One You’re

    Thinking of.” February 22nd, 2016. Written by Annie Gowen.

[51] Watson, Tara (2006). “Public Health Investments and the Infant Mortality Gap: Evidence from

    Federal Sanitation Interventions on U.S. Indian Reservations.” Journal of Public Economics 90:

    1537-1560.


                                                 27
[52] World Development Indicators (2017). “Mortality rate, infant (per 1000 live births)”.

    <http://data.worldbank.org/indicator/SP.DYN.IMRT.IN>. Accessed on 4/10/2017.

[53] World Health Organization (2004). “Global Health Observatory Data Repository.” Technical Re-

    port.




                                               28
              Figure 1: A district map of India




Notes: Light gray lines denote district borders, while black lines
denote state borders. The gray shaded area is the Ganga Basin
(240 districts), while the black shaded area is our group of four
“treatment” districts (Kanpur, Unnao, Fatehpur, and Rae Bareli).




                               29
           Figure 2: Pollution monitors in the Ganga basin




Notes: Black dots indicate the location of water pollution monitors. The
meandering dark gray line is the Ganga River; all monitors not on the Ganga
are situated along one of its tributaries (not shown). The gray shaded area
delineates the four “treatment” districts (Kanpur, Unnao, Fatehpur, and Rae
Bareli).




                                    30
Figure 3: Neonatal Mortality: Graphical Difference-in-Differences



                                                     Panel A. 1971-2004




                  1
                                                                               Treatment group
                                                                               Rest of Ganga Basin



                  .9
                  .8
    Annual neonatal mortality rate
    .3     .4   .5.2
                  .1
                  0    .6      .7




                                     1970   1975   1980   1985          1990   1995    2000      2005
                                                                 Year


                                                     Panel B. 1980-1994
                  .15




                                                                               Treatment group
                                                                               Rest of Ganga Basin
                  .13
    Annual neonatal mortality rate
         .09      .07
                  .05 .11




                                     1980   1982   1984   1986          1988   1990    1992      1994
                                                                 Year



  Notes: Both panels plot, separately for the treatment group and
  control group, annual averages of monthly neonatal mortality rates.
  Panel A displays averages from the full sample period of 1971-2000.
  Panel B focuses on the shorter time period of 1980-1994.




                                                            31
Figure 4: Biochemical Oxygen Demand: Graphical Difference-in-Differences



                                                                       Panel A. 1986-2004




                           1
                                                                                                           Treated
                                                                                                           Control




                                            .8
         Annual average of monthly 1[BOD>3]
                   .4      .2
                           0   .6




                                                 1986   1988   1990    1992   1994 1996   1998    2000   2002   2004
                                                                                 Year


                                                                       Panel B. 1986-1994
                           1




                                                                                                           Treated
                                                                                                           Control
                                            .8
         Annual average of monthly 1[BOD>3]
                   .4      .2
                           0   .6




                                                 1986   1987    1988     1989    1990   1991     1992    1993   1994
                                                                                 Year



      Notes: Both panels plot, separately for the treatment group and
      control group, annual averages of the monthly value of 1[BOD>3].
      Panel A displays averages from the full period of available data,
      1986-2004. Panel B focuses on the shorter time period of 1986-
      1994.




                                                                                32
                 Figure 5: Biological Oxygen Demand (BOD)



   2
                                         Treatment Districts
                                                                          Before
                                                                          After
   1.5
Density
  1.5
   0




          -2.5    -2      -1.5      -1       -.5    0      .5   1   1.5   2    2.5
                                             Residual ln(BOD)
          kernel = epanechnikov, bandwidth = 0.1326



                                          Control Districts
   2




                                                                          Before
                                                                          After
   1.5
Density
  1.5
   0




          -2.5    -2      -1.5      -1       -.5    0      .5   1   1.5   2    2.5
                                             Residual ln(BOD)
          kernel = epanechnikov, bandwidth = 0.1524


   Notes: Both panels show kernel densities of residuals from a re-
   gression of ln(BOD) on CETP capacity, air temperature, total pre-
   cipitation, 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.




                                                 33
                                                            Table 1: Summary Statistics for All Variables

     Panel A. Child-level variables                                         Treatment (Kanpur region)                          Control (Rest of Ganga Basin)

                                                                      N     Mean    St. Dev.   Min      Max                N       Mean    St. Dev.   Min      Max
     1[Child died in the first month of life]                        8,175    0.08     0.27      0         1             639,690     0.06     0.25      0         1
     1[Child died in the first year of life]                         8,175    0.11     0.32      0        1              639,690     0.09     0.29      0         1
     1[Child died in the first year of life | survived first month]   7,517    0.04     0.19      0         1             598,254     0.03     0.18      0        1
     1[Child was born after the verdict]                            8,175    0.81     0.40      0         1             639,690     0.83     0.38      0        1
     1[Mother is Hindu]                                             8,175    0.80     0.40      0         1             639,690     0.84     0.37      0        1
     1[Scheduled Caste/Scheduled Tribe]                             8,175    0.29     0.45      0         1             639,690    0.27      0.44      0        1
     Age of mother (years) at time of interview                     8,175   33.19     6.68     15        44             639,690    32.58     6.71     15        44
     1[Mother is literate]                                          8,175    0.35     0.48      0        1              639,690    0.33      0.47      0        1




34
     Panel B. District-level variables
                                                                      N     Mean    St. Dev.    Min      Max                N      Mean    St. Dev.   Min      Max
     Ln(BOD)                                                         383     1.07     0.27       0        2              8,048     0.94      1.06     -2         5
     1[BOD>3]                                                        383     0.34     0.48       0        1              8,048     0.39      0.49      0         1
     1[FCOLI>50]                                                     285     0.59     0.49       0        1               6,147     0.34     0.47      0         1
     1[Calc>median]                                                  380     0.36     0.48       0        1               6,394     0.51     0.50      0         1
     1[Sulfates>median]                                              294     0.50     0.50       0        1              6,275      0.49     0.50      0         1
     1[Chlorine>median]                                              386     0.59     0.49       0        1               6,842     0.49     0.50      0         1
     Air temperature (degrees C)                                    1,095   25.47     5.95     11.92    35.47            84,740    24.78     5.71     8.1      36.11
     Monthly precipitation (mm)                                     1,095   71.32    110.76      0      706.26           84,740    89.77    134.55     0     1,467.02
     1[National River Conservation Plan]                             398     0.56     0.50       0        1              8,337      0.43     0.50      0         1
     Common Effluent Treatment Plant capacity (MLD)                    398     0.00     0.00       0        0               8,337     0.12     1.29      0        24
     Notes: In Panel A, an observation is a child. In Panel B, it is a district-month. The sample marked ’Treatment’ consists of observations from Kanpur, Unnao,
     Fatehpur, and Rae Bareli districts. The sample marked ’Control’ consists of observations from all other districts in the Ganga Basin.
                               Table 2: Mehta vs. Union of India and Infant Mortality

                                           (1)          (2)           (3)           (4)           (5)          (6)
                                           IM           NM         Cond. IM         IM            NM        Cond. IM
     1[Kanpur] X 1[Post-Verdict]       -0.005∗∗∗     -0.024∗∗∗       -0.002      -0.004∗∗∗     -0.020∗∗∗       -0.001
                                        (0.001)       (0.006)       (0.001)       (0.001)       (0.005)       (0.001)
     Dependent variable mean             0.008         0.065         0.003        0.010         0.073         0.004
     Sample time frame                 1971-2004     1971-2004     1971-2004    1980-1994     1980-1994     1980-1994
     Adj. R-Squared                      0.003         0.014         0.001        0.003         0.017         0.001
     N                                 7,215,872      647,861      6,568,008    3,440,687      326,430      3,114,257




35
     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.
                                               Table 3: Timing and Scope of Neonatal Mortality Impacts

                                                                           Dependent variable: 1[Child died in first month of life]
                                                              (1)           (2)           (3)           (4)           (5)          (6)           (7)
     1[Kanpur] X 1[Post-Verdict]                           -0.026∗∗∗     -0.027∗∗∗     -0.020∗∗∗                  -0.026∗∗∗                   -0.020∗∗∗
                                                            (0.006)       (0.005)       (0.005)                    (0.005)                     (0.005)
     1[Kanpur] X 1[10/1987 <= t <= 12/1994]                                                          -0.020∗∗∗
                                                                                                      (0.005)
     1[Kanpur] X 1[1/1995 <= t <= 12/1999]                                                           -0.035∗∗∗
                                                                                                      (0.006)
     1[Kanpur] X 1[1/2000 <= t <= 12/2004]                                                           -0.024∗∗
                                                                                                      (0.010)
     Kanpur X 1[Some in utero treatment]                                                                                          0.016
                                                                                                                                 (0.015)
     Kanpur X 1[Full in utero treatment]                                                                                        -0.029∗∗∗
                                                                                                                                 (0.010)




36
     1[Near Downstream] X 1[Post-Verdict]                                                                                                     -0.032∗∗∗
                                                                                                                                               (0.005)
     1[Intermediate Downstream] X 1[Post-Verdict]                                                                                               0.000
                                                                                                                                               (0.007)
     1[Far Downstream] X 1[Post-Verdict]                                                                                                        -0.004
                                                                                                                                               (0.007)
     Dependent variable mean                                 0.064        0.067         0.073         0.064         0.059         0.073        0.073
     Geographic coverage                                    Ganga        Ganga         Ganga         Ganga          India        Ganga        Ganga
     Sample time frame                                     1980-2004    1980-1999     1980-1994     1980-2004     1980-1994     1980-1994    1980-1994
     Adj. R-Squared                                          0.013        0.014         0.017         0.013         0.021         0.017        0.017
     N                                                      633,530      509,490       326,430       633,530       720,640       326,430      326,430
     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.
     ***, **, and * indicate statistical significance at the 1, 5, and 10 percent levels, respectively.
                                                            Table 4: Mehta vs. Union of India and Pollution

     Panel A. Biochemical Oxygen Demand                                                        1[BOD>3]                                                     Ln(BOD)
                                                               (1)            (2)            (3)            (4)         (5)            (6)            (7)             (8)
                                                                     ∗∗∗            ∗∗∗            ∗∗∗                        ∗∗∗            ∗∗∗            ∗
     1[Kanpur] X 1[Post-Verdict]                            -0.534         -0.568         -0.607                     -0.696         -0.637          -0.236        -0.108
                                                             (0.058)        (0.050)        (0.055)                    (0.028)        (0.052)        (0.125)      (0.116)
     1[Kanpur] X 1[10/1987 <= t <= 12/1994]                                                              -0.492∗∗∗
                                                                                                          (0.133)
     1[Kanpur] X 1[1/1995 <= t <= 12/1999]                                                               -0.548∗∗∗
                                                                                                          (0.079)
     1[Kanpur] X 1[1/2000 <= t]                                                                          -0.578∗∗∗
                                                                                                          (0.133)
     1[Near Downstream X 1[Post-Verdict]                                                                                            -0.655∗∗∗
                                                                                                                                     (0.051)
     1[Intermediate Downstream X 1[Post-Verdict]                                                                                      -0.249
                                                                                                                                     (0.229)
     1[Far Downstream X 1[Post-Verdict]                                                                                               0.060




37
                                                                                                                                     (0.058)
     Dependent variable mean                                 0.392           0.408          0.416          0.392        0.286         0.416          0.947        0.889
     Geographic coverage                                    Ganga           Ganga          Ganga          Ganga         India        Ganga          Ganga        Ganga
     Sample Time Frame                                     1986-2004       1986-2000      1986-1994      1986-2004   1986-1994      1986-1994      1986-2004    1986-1994
     Adj. R-Squared                                          0.563           0.559          0.594          0.563        0.553         0.608          0.781        0.800
     N                                                       8,431           5,530          2,883          8,431       12,521         2,883          8,431        2,883
     Panel B. Non-BOD Pollutants                            Calcium        Sulfates       Chlorides       FCOLI
                                                               (1)            (2)            (3)            (4)
     1[Kanpur] X 1[Post-Verdict]                            -0.178∗∗       -0.444∗∗∗      -0.149∗∗         -0.010
                                                             (0.074)        (0.048)        (0.056)        (0.101)
     Dependent variable mean                                 0.631           0.551          0.458          0.447
     Geographic coverage                                    Ganga           Ganga          Ganga          Ganga
     Sample Time Frame                                     1986-1994       1986-1994      1986-1994      1986-1994
     Adj. R-Squared                                          0.595           0.412          0.650          0.531
     N                                                       2,616           2,542          3,018          2,003
     Notes: All results are based on estimation of equation (2). An observation is a district-month. The dependent variable is listed above each column number.
     All regressions include a set of controls (CETP capacity, air temperature, total precipitation, and NRCP dummy) and district and year-month fixed effects.
     Standard errors are clustered at the district level in parentheses. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent levels, respectively.
                                  Table 5: Impact of Upstream Pollution on Downstream Mortality

                                                              Dependent variable: 1[Died in first month]
                                      OLS                       Upstream Districts                               Placebo Districts
                                       (1)           (2)          (3)           (4)           (5)                 (6)           (7)
     1[BOD>3]                         -0.007
                                     (0.010)
     1[US BOD>3]                                  0.030∗∗∗      0.020∗∗      0.022∗∗∗       0.030∗∗              -0.006       0.006
                                                   (0.010)      (0.009)       (0.007)       (0.014)             (0.009)      (0.004)
     Dependent variable mean          0.067         0.067         0.058        0.067         0.066              0.071          0.057
     Upstream definition             [75,200]      [75,200]      [75,200]     [0,200]       [100,200]            N/A            N/A




38
     Last sample year              1986-1994     1986-1994     1986-2004    1986-1994     1986-1994           1986-1994     1986-2004
     Adj. R-Squared                  0.015         0.016         0.010        0.016          0.016              0.012         0.011
     N                                9,603         9,603        32,560       13,608         8,979              2,626         16,891
     Notes: An observation is a child-month. The dependent variable in all regressions is a binary variable equaling one if a child
     died in the first month of life. Column 1 shows the results from OLS estimation of mortality on pollution. Columns 2-5 show
     results from estimation of equation (3). Columns 6 and 7 show results from a placebo test, in which the upstream district for
     each observation is replaced by a different, neighboring district that is not upstream. In each column, the sample is defined by
     its listed ’Upstream definition’, ’Geographic coverage’, and ’Last sample year’ ("N/A" indicates no sample restriction based on
     upstream distance). All regressions additionally 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.
                                                  Table 6: Instrumental Variables Impact of Pollution on Mortality

     Panel A. 2SLS first stage                                                          1-IV                                                        2-IV
                                                              (1)           (2)                 (3)           (4)           (5)         (6)                 (7)           (8)
                                                                 ∗∗∗              ∗∗                  ∗∗            ∗∗        ∗∗∗             ∗∗                  ∗∗
     1[US BOD>3]                                           0.211          0.167               0.180         0.172        0.211        0.167               0.180         0.172∗∗
                                                            (0.064)       (0.065)             (0.073)       (0.072)       (0.064)     (0.065)             (0.073)       (0.072)
     1[Kanpur] X 1[Post-Verdict]                           -0.571∗∗∗     -0.462∗∗∗        -0.389∗∗∗                      -0.571∗∗∗   -0.462∗∗∗        -0.389∗∗∗
                                                            (0.068)       (0.058)          (0.075)                        (0.068)     (0.058)          (0.075)
     1[Kanpur] X 1[10/1987 <= t <= 12/1994]                                                                -0.446∗∗∗                                                   -0.446∗∗∗
                                                                                                            (0.143)                                                     (0.143)
     1[Kanpur] X 1[1/1995 <= t <= 12/1999]                                                                 -0.353∗∗∗                                                   -0.353∗∗∗
                                                                                                            (0.102)                                                     (0.102)
     1[Kanpur] X 1[1/2000 <= t]                                                                             -0.316∗                                                     -0.316∗
                                                                                                            (0.189)                                                     (0.189)
     Panel B. 2SLS second stage                               (1)           (2)                 (3)           (4)           (5)         (6)                 (7)           (8)




39
                                                                    ∗∗            ∗∗               ∗∗∗          ∗∗∗           ∗∗∗         ∗∗∗                  ∗∗∗
     1[BOD>3]                                               0.144         0.126               0.111        0.121         0.090       0.100                0.102        0.107∗∗∗
                                                            (0.061)       (0.054)              (0.034)      (0.040)       (0.020)     (0.032)              (0.027)      (0.025)
     1[Kanpur] X 1[Post-Verdict]                             0.050         0.018               0.009
                                                            (0.043)       (0.028)             (0.019)
     Dependent variable mean                                 0.067         0.61                0.058        0.058          0.067       0.061            0.058            0.058
     F-statistic (p-value)                                   <0.01         <0.01               <0.01        <0.01          <0.01       <0.01            <0.01            <0.01
     C-statistic (p-value)                                                                                                 0.348       0.525            0.634             0.469
     Sample time period                                   1986-1994      1986-1999        1986-2004        1986-2004     1986-1994   1986-1999        1986-2004        1986-2004
     N                                                      9,603          21,703           32,561           32,561        9,603       21,703           32,561           32,561
     Notes: An observation is a child-month. The dependent variable in all regressions is a binary variable equaling one if a child died in the first month of life. All
     columns display results from Two-Stage Least Squares estimation of equations (4) and (5); panel A details 2nd-stage results and panel B details 1st-stage results.
     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.
A   Appendix tables and figures

         Figure A1: Time trends in the treatment group vs. Kannauj district




                                       .3
                                                                                            Treatment group              Kannauj




                    Annual average of monthly mortality rate
                           .1          0         .2




                                                               1980          1985       1990          1995       2000          2005
                                                                                               Year
                                       1




                                                                                            Treatment group              Kannauj
                                       .9              .8
                    Annual average of monthly 1[BOD>3]
                       .3     .4    .5 .2
                                       .1
                                       0   .6     .7




                                                               1986   1988    1990   1992   1994 1996    1998   2000    2002   2004
                                                                                               Year



                  Notes: Panel A plots annual average mortality; Panel
                  B plots annual average BOD exceedance (1[BOD>3]).
                  Both panels plot averages in the treatment group and
                  control group separately. The treatment group is com-
                  prised of Kanpur Nagar, Unnao, Fatehpur, and Rae
                  Bareli districts. The control group includes only a
                  single upstream district, Kannauj.




                                                                                            40
Figure A2: Time trends in Kanpur Nagar district vs. the control group


                                                                     Panel A. 1980-2004




                         .2
                                                                                             Kanpur Nagar
                                                                                             Rest of Ganga Basin




                                      .15
       Annual average of monthly mortality
       .05            .1 0




                                               1980          1985       1990          1995        2000          2005
                                                                               Year


                                                                     Panel B. 1986-2004
                         1




                                                                                                   Kanpur Nagar
                                                                                                   Control
                         .9               .8
       Annual average of monthly 1[BOD>3]
          .3     .4    .5.2
                         .1
                         0   .6     .7




                                               1986   1988    1990   1992   1994 1996    1998   2000     2002   2004
                                                                               Year



     Notes: Panel A plots annual average mortality; Panel B plots an-
     nual average BOD exceedance (1[BOD>3]). Both panels plot av-
     erages in the treatment group and control group separately. The
     treatment group is Kanpur Nagar district only. The control group
     is the rest of the Ganga Basin.




                                                                             41
                                                      Table A1: Pre-Ruling Summary Statistics for All Variables

     Panel A. Child-level variables                                          Treatment (Kanpur region)                        Control (Rest of Ganga Basin)

                                                                       N       Mean    St. Dev.   Min    Max                N       Mean    St. Dev.    Min    Max
     1[Child died in the first month of life]                         1,821      0.12     0.32      0       1             125,931     0.09     0.29       0       1
     1[Child died in the first year of life]                          1,821      0.17     0.38      0      1              125,931     0.13     0.34       0       1
     1[Child died in the first year of life | survived first month]    1,582      0.06     0.24      0      1              112,623     0.04     0.21       0      1
     1[Mother is Hindu]                                              1,821      0.79     0.40      0       1             125,931     0.85     0.36       0      1
     1[Scheduled Caste/Scheduled Tribe]                              1,821      0.25     0.43      0      1              125,931     0.26     0.44       0      1
     Age of mother (years) at time of interview                      1,821     39.54     3.34     28      44             125,931    39.55     3.26      21      44
     1[Mother is literate]                                           1,821      0.33     0.47      0      1              125,931     0.30     0.46       0      1
     Panel B. District-level variables




42
                                                                      N        Mean    St. Dev.   Min    Max                N       Mean    St. Dev.    Min    Max
     Ln(BOD)                                                          22        1.66     0.31      1      2                468       1.09     0.85       -1     3
     1[BOD>3]                                                         22        1.00     0.00      1      1                468       0.46     0.50       0      1
     1[FCOLI>50]                                                      22        1.00     0.00      1      1                411       0.51     0.50       0      1
     1[Calc>median]                                                   22        0.86     0.35      0      1                367       0.47     0.50       0      1
     1[Sulfates>median]                                               22        0.91     0.29      0      1                404       0.47     0.50        0      1
     1[Chlorine>median]                                               22        0.50     0.51      0      1                435       0.50     0.50       0      1
     Air temperature (degrees C)                                      22       26.27     5.90     16      35               471      26.33     5.24      11      36
     Monthly precipitation (mm)                                       22       53.11    75.98      0     291               471      95.52    135.10      0     753
     1[National River Conservation Plan]                              22       1.00      0.00      1      1                471      0.42      0.49       0       1
     Common Effluent Treatment Plant capacity (MLD)                     22        0.00     0.00      0      0                471       0.00     0.00       0      0
     Notes: All statistics are based on pre-Supreme Court verdicts data. In Panel A, an observation is a child. In Panel B, it is a district-month. The sample marked
     ’Treatment’ consists of observations from Kanpur, Unnao, Fatehpur, and Rae Bareli districts. The sample marked ’Control’ consists of observations from all other
     districts in the Ganga Basin.
       Table A2: Sample Composition and Upstream Definition

                  Ganga Basin Only                     All India

             Number of       Upstream        Number of       Upstream
             Monitors      Distance (km)     Monitors      Distance (km)
[0, 200]        77               96            343               87
[20, 200]       70              102            333               92
[50, 200]       68              114            315              108
[75, 200]       54              128            278              124
[100, 200]      48              146            255              141
[75, 300]       78              177            328              149
Notes: The window [X,Y] defines the range, in km, of distances at which a
pollution monitor lying upstream of some monitor m qualifies as its upstream
match. That match is also conditional on the upstream monitor lying in a
different district than monitor m. Tabulated numbers count the monitors
in the sample that have upstream matches, as well as the average distance
upstream of those matches. In the left panel, the sample is the Ganga Basin
only; in the right, it is all of India




                                    43
                                   Table A3: Summary Statistics for Variables in the Merged Mortality-Pollution Sample

                                                                       Treatment (Kanpur region)                          Control (Rest of Ganga Basin)

                                                                 N       Mean    St. Dev.    Min    Max                N       Mean     St. Dev.    Min    Max
     1[Child died in the first month of life]                   2,962      0.06     0.24       0       1              29,599     0.06      0.23       0       1
     1[BOD>3]                                                  2,962      0.38     0.48       0       1              29,599     0.41      0.49       0       1
     Upstream 1[BOD>3]                                         2,962      0.92     0.28       0       1              29,599    0.36       0.48       0       1
     1[Child was born after the verdict]                       2,962      0.95     0.21       0       1              29,599     0.96      0.20       0       1
     1[Mother is Hindu]                                        2,962      0.73     0.45       0       1              29,599     0.80      0.40       0       1




44
     1[Scheduled Caste/Scheduled Tribe]                        2,962      0.26     0.44       0       1              29,599     0.27      0.44       0       1
     Age of mother (years) at time of interview                2,962     31.27     6.11      16      44              29,599    30.15      6.35      15      44
     1[Mother is literate]                                     2,962     0.43      0.49       0       1              29,599     0.37      0.48       0      1
     Monthly mean of daily mean airtemp (C)                    2,962     25.93     5.80      14      35              29,599    25.45      5.69      10      36
     Interpolated Precipitation (mm)                           2,962     73.21    98.83       0     406              29,599    85.46     121.67      0     742
     1[National River Conservation Plan]                       2,962      0.53     0.50       0       1              29,599     0.49      0.50       0       1
     Common Effluent Treatment Plant capacity (MLD)              2,962      0.00     0.00       0       0              29,599     0.23      1.18       0       6
     Notes: The sample includes all child-months with non-missing values of the variables listed, which are those required for IV estimation. The upstream buffer
     used is [75 km, 200 km]. “N” lists the number of child-month observations. The sample marked ’Treatment’ consists of observations from Kanpur, Unnao,
     and Rae Bareli districts. The sample marked ’Control’ consists of observations from all other districts in the Ganga Basin.
                    Table A4: Policy Impact Regressions with Merged Mortality-Pollution Sample

                                           NM            NM            NM          1[BOD>3]       1[BOD>3]       1[BOD>3]
                                           (1)           (2)            (3)           (4)            (5)            (6)
                                                 ∗∗            ∗∗∗            ∗∗            ∗∗∗            ∗∗∗
     1[Kanpur] X 1[Post-Verdict]        -0.033        -0.040         -0.034        -0.381         -0.455         -0.587∗∗∗
                                         (0.013)       (0.013)        (0.014)       (0.076)        (0.059)        (0.071)
     Dependent variable mean             0.058          0.061          0.067         0.408          0.391          0.367
     Sample Time Frame                 1986-2004      1986-1999      1986-1994     1986-2004      1986-1999      1986-1994
     Adj. R-Squared                      0.009          0.013          0.015         0.681          0.646          0.686




45
     N                                   32,560         21,703         9,603         32,560         21,703         9,603
     Notes: The sample includes all child-months with non-missing values for neonatal survival, BOD, and upstream
     BOD (with ‘upstream’ defined as being in the range [75 km, 200 km]). Columns 1-3 correspond to estimation of
     equation (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.
                                   Table A5: Robustness Checks on Instrumental Variables Regressions

     Panel A. 2SLS first stage                                     1-IV                                                   2-IV
                                             (1)         (2)               (3)           (4)         (5)         (6)            (7)           (8)
                                                ∗∗∗         ∗∗∗                  ∗∗        ∗∗∗         ∗∗∗         ∗∗∗                ∗∗
     1[US BOD>3]                          0.169        0.213         0.207            0.275       0.169       0.213         0.207          0.275∗∗∗
                                           (0.063)      (0.062)      (0.093)           (0.039)     (0.063)     (0.062)      (0.093)         (0.039)
     1[Kanpur] X 1[Post-Verdict]          -0.561∗∗∗   -0.604∗∗∗      -0.459∗∗∗        -0.674∗∗∗   -0.561∗∗∗   -0.604∗∗∗     -0.459∗∗∗      -0.674∗∗∗
                                           (0.072)     (0.061)        (0.092)          (0.039)     (0.072)     (0.061)       (0.092)        (0.039)
     Panel B. 2SLS second stage              (1)         (2)               (3)           (4)         (5)         (6)            (7)           (8)
     1[BOD>3]                             0.154∗∗      0.110∗∗            0.137        0.028      0.096∗∗∗    0.074∗∗∗      0.139∗∗∗       0.035∗
                                          (0.062)      (0.044)           (0.085)      (0.020)      (0.027)     (0.018)       (0.034)       (0.019)
     1[Kanpur] X 1[Post-Verdict]            0.054       0.035             -0.001      -0.031∗∗
                                           (0.042)     (0.034)           (0.043)       (0.015)




46
     Dependent variable mean               0.065        0.066            0.067         0.053        0.065       0.066        0.067           0.053
     F-statistic (p-value)                 <0.01        <0.01            <0.01         <0.01       <0.01       <0.01         <0.01          <0.01
     C-statistic (p-value)                                                                         0.259        0.393        0.978           0.214
     Upstream Range (km)                  [20,200]     [50,200]      [100,200]        [75,200]    [20,200]    [50,200]      [50,200]       [75,200]
     Geographic coverage                   Ganga        Ganga         Ganga             India      Ganga       Ganga         Ganga           India
     N                                     13,309       11,573         8,509           60,511      13,309      11,573         8,509         60,511
     Notes: An observation is a child-month. The dependent variable in all regressions is a binary variable equaling one if a child died in
     the first month of life. All columns display results from Two-Stage Least Squares estimation of equations (4) and (5); panel A details
     2nd-stage results and panel B details 1st-stage results. 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.
B     A model to test the mechanisms of impact

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 ortalityidt = a + bTdt + Xidt γ + eidt                            (A.1)


    In this reduced-form model of infant mortality, identification rests on the assumption that Cov Tdt , eidt =

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:


                                                           ˜ idt γ + (Zidt δ + εidt )
                       M ortalityidt = α + βP ollutiondt + X                                         (A.2)


    This equation is analogous to equation (5) in the main text, except that it partitions the space

of unobserved risk factors into two: Zidt and εidt . The former is a vector of all factors that are

also correlated with environmental policy Tdt . These include, but are not restricted to, individual

awareness about river water contamination, changes in factor prices stemming from the implementation

of environmental policy Tdt , or any type of private or public interventions that might have been

triggered by Tdt . The latter captures the other risk factors of infant mortality and is, by construction,
                           ˜ idt = 0.
such that Cov Tdt , εidt | X

    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 Zidt and P ollutiondt respond to

environmental policy and other determinants in linear fashion. That is,


                                                         ˜ idt γ 1 + ε1
                                   Zidt = α1 + β 1 Tdt + X            idt                            (A.3)


and
                                                             ˜ idt γ 2 + ε2
                               P ollutiondt = α2 + β 2 Tdt + X                                       (A.4)
                                                                          idt


We can then rewrite equation (A.2) by substituting for both Zidt and P ollutiondt , so as to decompose




                                                    47
the policy’s effect on mortality into the different channels of impact:


                   M ortalityidt   =    α + βα2 + α1 δ + ββ 2 + β 1 δ Tdt

                                         ˜ idt βγ 2 + γ + γ1 δ + βε2
                                        +X                                1
                                                                   idt + εidt δ + εidt               (A.5)


The total impact of environmental policy Tdt 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 Zidt . This yields


     M ortalityidt =                               ˜ idt γ + γ 1 δ + β 1 δ Tdt + ε1 δ + εidt
                        α + α1 δ + βP ollutiondt + X                                                 (A.6)
                                                                                  idt



which contains both β and the product β 1 δ . We estimate equation (A.6) using 2SLS, where we

instrument for P ollutiondt with its upstream analog, P ollutionu
                                                                dt . With unbiased estimates of both

β and β 1 δ , we can then test H0 : 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


                                                        ˜ idt γ + γ 1 · δ + ε1 δ + εidt
             M ortalityidt = α + α1 δ + βP ollutiondt + X                                            (A.7)
                                                                             idt



so that Tdt is additionally excluded from second-stage equation (A.7) and becomes another valid in-

strument for P ollutiondt . One test of H0 is therefore an over-identification test that assesses the

orthogonality condition for Tdt , as part of the larger set of instruments {Tdt , P ollutionu
                                                                                            dt }. To imple-

ment such a test, we construct a C-statistic (see, e.g., Eichenbaum, Hansen, and Singleton 1988),which

we describe further in the main text.




                                                    48