Consequences of the Clean Water Act and the Demand for Water Quality*

Abstract

I. Introduction

The 1972 U.S. Clean Water Act sought “to restore and maintain the chemical, physical, and biological integrity of the Nation’s waters.” This article quantifies changes in water pollution since before 1972, studies the causes of any changes, and analyzes the welfare consequences of any changes.

The Clean Water Act addressed a classic externality. Textbooks since at least Stigler (1952, 1966) have illustrated the concept of an externality through the story of a plant dumping waste in a river and harming people downstream. The immediate impetus for the Clean Water Act was a 1969 fire on the Cuyahoga River, which had fires every decade since 1868 but has had no fires since 1972. Time (1969) described it vividly:

Anyone who falls into the Cuyahoga does not drown,” Cleveland’s citizens joke grimly. “He decays.” The Federal Water Pollution Control Administration dryly notes: “The lower Cuyahoga has no visible life, not even low forms such as leeches and sludge worms that usually thrive on wastes. It is also literally a fire hazard.

Despite the potential to address this market failure, the Clean Water Act has been one of the most controversial regulations in U.S. history for at least two reasons. First, it is unclear whether the Clean Water Act has been effective or whether water pollution has decreased at all. An analysis in the 1990s summarized, “As we approached the twenty-year anniversary of this landmark law, no comprehensive analysis was available to answer basic questions: How much cleaner are our rivers than they were two decades ago?” (Adler, Landman, and Cameron 1993). Other writers echo these sentiments (Knopman and Smith 1993; Powell 1995; Harrington 2004). Today data indicate that over half of U.S. river and stream miles violate state water quality standards (USEPA 2016), but it is not known if water quality was even worse before the Clean Water Act. William Ruckelshaus, the first head of the U.S. Environmental Protection Agency (EPA), nicely summarized what is known about water pollution today: “even if all of our waters are not swimmable or fishable, at least they are not flammable” (Mehan 2010).

The second controversy is whether the Clean Water Act’s benefits have exceeded its costs, which have been enormous. Since 1972, government and industry have spent over |${\$}$|1 trillion to abate water pollution, or over |${\$}$|100 per person-year. This is more than the United States has spent on air pollution abatement (see Online Appendix A). In the mid-1970s, Clean Water Act funding of municipal wastewater treatment plants was the single largest public works program in the United States (USEPA 1975). These costs were large partly because the Clean Water Act had ambitious targets: to make all U.S. waters fishable and swimmable by 1983; to have zero water pollution discharge by 1985; and to prohibit discharge of toxic amounts of toxic pollutants. President Richard Nixon actually vetoed the Clean Water Act and described its costs as “unconscionable,” though Congress later overruled the veto (Nixon 1972). Large costs could be outweighed by large benefits. However, existing cost-benefit analyses of the Clean Water Act have not estimated positive benefit/cost ratios (Lyon and Farrow 1995; Freeman 2000; Keiser, Kling, and Shapiro forthcoming), including the EPA’s own retrospective analysis (USEPA 2000a, b).

These academic controversies have spilled over into politics. The U.S. Supreme Court’s 2001 and 2006 SWANCC and Rapanos decisions removed Clean Water Act regulation for nearly half of U.S. rivers and streams. In 2015, the Obama administration proposed a Clean Water Rule (also called the Waters of the United States Rule) that would reinstate many of those regulations. Twenty-seven states have sued to vacate the rule.

This article seeks to shed light on these controversies using the most comprehensive set of files ever compiled in academia or government on water pollution and its determinants. These files include several data sets that largely have not been used in economic research, including the National Hydrography Dataset, which is a georeferenced atlas mapping all U.S. surface waters; the Clean Watershed Needs Survey, a panel description of the country’s wastewater treatment plants; a historic extract of the Grants Information and Control System describing each of 35,000 Clean Water Act grants the federal government gave to cities; the Survey of Water Use in Manufacturing, a confidential plant-level data set of large industrial water users that was recently recovered from a decommissioned government mainframe (Becker 2016); and around 50 million water pollution readings at over 240,000 pollution monitoring sites during the years 1962–2001 from three data repositories—Storet Legacy, Modern Storet, and the National Water Information System (NWIS). Discovering, obtaining, and compiling these data has been a serious undertaking involving three Freedom of Information Act requests, detailed analysis of hydrological routing models, and extensive discussions with engineers and hydrologists from the U.S. Geological Survey (USGS), the EPA, and engineering consultancies. These data enable a more extensive analysis of water pollution and its regulation than has previously been possible.

The analysis obtains three sets of results. First, we find that most types of water pollution declined over the period 1962–2001, though the rate of decrease slowed over time. Between 1972 and 2001, the share of waters that met standards for fishing grew by 12 percentage points. The pH of rivers and lakes has increased at a similar rate to the pH of rainwater, likely in part due to decreased sulfur air pollution. In other words, less acid rain may have led to less acidic rivers and lakes. In addition, the temperature of rivers and lakes increased by 1°F every 40 years, consistent with climate change.

Second, the article asks how the Clean Water Act’s grants to municipal wastewater treatment plants, one of the act’s central components, contributed to these trends. We answer this question using a triple-difference research design comparing water pollution before versus after investments occurred, upstream versus downstream of recipient plants, and across plants. We define upstream and downstream waters using a set of 70 million nodes that collectively describe the entire U.S. river network. We find that each grant decreases the probability that downriver areas violate standards for being fishable by half a percentage point. These changes are concentrated within 25 miles downstream of the treatment plant and they persist for 30 years. Through these grants, it cost around |${\$}$|1.5 million (⁠|${\$}$|2014) a year to make one river-mile fishable. We do not find substantial heterogeneity in cost-effectiveness across regions or types of grants. We also find that |${\$}$|1 of a federal grant project led to about |${\$}$|1 more of municipal sewerage capital spending.

Third, the article asks how residents valued these grants. We analyze housing units within a 25-mile radius of affected river segments, partly since 95% of recreational trips have a destination within this distance. We find that a grant’s estimated effects on home values are about 25% of the grant’s costs. While the average grant project in our analysis cost around |${\$}$|31 million (⁠|${\$}$|2014), our main estimates imply that the estimated effect of a grant on the value of housing within 25 miles of the affected river is around |${\$}$|7 million. We find limited heterogeneity in these numbers across regions and types of grants.

We discuss several reasons the true benefit/cost ratio for the grants program could exceed this 0.25 ratio of the change in home values to grant costs. These reasons include that people may have incomplete information about changes in water pollution and their welfare (including health) implications; these numbers exclude nonuse (“existence”) values; grants may increase sewer fees; these estimates abstract from general equilibrium effects; and they exclude the 5% of most distant recreational trips. Available evidence to evaluate these reasons is limited; it does suggest that the true benefit/cost ratio may exceed 0.25, but does not clearly show that this ratio exceeds 1. One interpretation of the main estimates is that the benefits of these grants exceed their costs if these unmeasured components of willingness to pay exceed the components of willingness to pay that we measure by a factor of 3 or more.

We provide several indirect tests of the identifying assumptions, which generally support the validity of the research design. First, we report event study graphs in time, which test for pretrends in the years preceding a grant. Second, we report two research designs—a triple-difference estimator that uses upstream areas as a counterfactual for downstream areas, and a difference-in-differences estimate using only downstream areas. Third, we assess whether grants affect pollutants closely related to municipal waste more than they affect pollutants that are less closely related. Fourth, we separately estimate the effect of a plant receiving one, two, three, or more grants. Finally, we estimate specifications controlling for important potential confounding variables, including industrial water pollution sources, air pollution regulations, and local population totals.

More broadly, this article departs from the literature in four primary ways. This is the first study quantifying national changes in water pollution since before the Clean Water Act using a dense network of monitoring sites. Trends are important both in their own right and because measuring water pollution is a step toward measuring its costs (Muller, Mendelsohn, and Nordhaus 2011). Some studies measure trends in water pollution for limited sets of monitoring sites (e.g., Smith, Alexander, and Wolman 1987; USEPA 2000c).¹

This article also provides the first national estimate of how Clean Water Act investments affected ambient pollution concentrations. We use these estimates to calculate the cost-effectiveness of these investments. Water pollution research typically uses ex ante engineering simulations to assess water quality policies (Wu et al. 2004). A few studies do investigate how water pollution affects self-reported emissions of one pollutant in specific settings (Earnhart 2004a,b; Cohen and Keiser 2017), or study similar questions for air pollution (Shapiro and Walker forthcoming). Recent research finds that India’s water pollution regulations, which have a similar structure to the U.S. Clean Water Act, are ineffective (Ebenstein 2012; Greenstone and Hanna 2014). Several studies find that ambient water pollution increases with political boundaries (Sigman 2002; Kahn, Li, and Zhao 2015; Lipscomb and Mobarak 2017). Some work investigates how fracking wells and the pollution they send to wastewater treatment plants affect water quality (Olmstead et al. 2013).

Third, this study provides the first estimate of the effects of water pollution regulation on home values. Existing estimates of willingness to pay for water quality use travel cost methods, hedonics, or stated preferences (i.e., contingent valuation; Kuwayama and Olmstead 2015 list many individual studies).² Travel cost studies typically rely on cross-sectional variation in pollution and focus on a limited area like a county (e.g., Smith and Desvousges 1986), though some work uses broader coverage (Keiser forthcoming). Such studies may suffer from omitted variable bias because unobserved disamenities like factories or roads contribute to pollution and discourage recreational visits (Leggett and Bockstael 2000; Murdock 2006; Moeltner and von Haefen 2011). Such omitted variables are important for studying air pollution, though their importance for water pollution is unknown. Most cost-benefit analyses of the Clean Water Act rely on stated preferences (Carson and Mitchell 1993; Lyon and Farrow 1995; USEPA 2000a), which are controversial (Hausman 2012; Kling, Phaneuf, and Zhao 2012; McFadden and Train 2017).

Finally, we believe this is the first empirical study of the efficiency of subsidizing the use of pollution control equipment. This policy is common in many countries and settings. Theoretical research has lamented the poor incentives of such subsidies (Kohn 1992; Aidt 1998; Fredriksson 1998) and empirical research is scarce.³ Our analysis of heterogeneity in cost-effectiveness and benefit-cost ratios also provides a new domain to consider recent research on spatially differentiated policy (Muller and Mendelsohn 2009).

The article proceeds as follows. Section II describes the Clean Water Act and water pollution. Section III explains the data. Section IV discusses the econometric and economic models. Section V summarizes pollution trends. Section VI analyzes how grants affected pollution. Section VII discusses grants’ effects on housing. Section VIII concludes. All appendix material appears in the Online Appendix.

II. The Clean Water Act and Water Pollution

II.A. Clean Water Act Background

Policies before the Clean Water Act may contribute to some of the water pollution patterns we observe before 1972.⁴ The U.S. Congress passed major water pollution control laws in 1948, 1956, 1961, 1965, 1966, and 1970. Many earlier laws, like the Clean Water Act, supported municipal wastewater treatment and industrial abatement, but provided funds an order of magnitude below the funds distributed by the Clean Water Act. By 1966, all 50 states had passed some type of water pollution legislation, but enforcement varied greatly (Hines 1967).

The Clean Water Act retained large roles for state-level implementation, and the effectiveness of that implementation most likely varied across states. While a simple formula determined the level of grant funds that each state received, each state designed the priority lists determining which plants received grants. States with decentralized authority also oversaw writing of permits for municipal plants, monitoring and enforcement of violations, and other activities (Sigman 2003, 2005).

The Clean Water Act targeted municipal waste treatment and industrial pollution sources, sometimes called “point sources.” However, much water pollution also comes from “nonpoint” pollution sources such as urban and agricultural runoff. The Clean Water Act has largely exempted these latter sources from regulation.

This article focuses on the Clean Water Act grants program, but the act also limited industrial water pollution through the National Pollutant Discharge Elimination System (NPDES). NPDES aims to cover every source that directly discharges pollution into U.S. waters. Some plants are part of a separate “Pretreatment Program,” in which they discharge untreated or lightly treated wastewater through sewers to wastewater treatment plants, then pay fees to the treatment plant.⁵ The permits were distributed in the early 1970s. This was a national program affecting most plants and industries at around the same time.

II.B. Wastewater Treatment Background

In most cities and towns, sewers convey wastewater to a municipal wastewater treatment plant, which treats the waste and then discharges it to surface waters. Ninety-eight percent of treatment plants are publicly owned (USEPA 2002). The abatement technology in treatment plants initially only included screens to remove large objects. As technology improved during the twentieth century, treatment plants began allowing wastewater to settle before discharging, then began applying biological treatments (e.g., bacteria) that degrade pollution, and finally began using more advanced chemical treatments. These abatement technologies are generally called raw, primary, secondary, and tertiary treatment. The Clean Water Act required all municipal treatment plants to have at least secondary treatment by 1977.

This investment in wastewater treatment was not cheap. Projects funded by Clean Water Act grants cost about |${\$}$|650 billion in total over their lifetimes (⁠|${\$}$|2014). Grants covered new treatment plants, improvement of existing plants, and upgrades to sewers (USEPA 1975). Local governments paid about a fourth of most grant projects’ capital costs.⁶ The 1987 Clean Water Act Amendments replaced these grants with subsidized loans (the Clean Water State Revolving Fund).

The United States did not come close to meeting the Clean Water Act’s goal of having every plant install secondary treatment by 1977, though abatement technologies improved over time. In 1978, for example, nearly a third of all plants lacked secondary treatment, and by 1996, almost none did. The treatment technology used in wastewater treatment plants, however, had been improving steadily before the Clean Water Act (USEPA 2000c).

Because this article exploits the timing and location of grants to identify the effect of the Clean Water Act’s grants program, it is useful to clarify how grants were distributed. The allocation of wastewater spending across states came from formulas depending on state population, forecast population, and wastewater treatment needs (CBO 1985). Within a state, grants were distributed according to a “priority list” that each state submitted annually to the EPA. States had to base a priority list on seven criteria:

1. [T]he severity of the pollution problem; 2. [T]he existing population affected; 3. [T]he need for preservation of high quality waters; 4. [A]t the State’s option, the specific category of need … 5. … [T]echniques meeting innovative and alternative guidelines … 6. [O]ther criteria, consistent with these, may be considered (including the special needs of small and rural communities). The state may not consider: the project area’s development needs not related to pollution abatement; the geographical region within the State; or future population growth projections; and 7. [I]n addition to the criteria listed above, the State must consider … total funds available; and other management criteria. (USEPA 1980, 8)

The EPA estimated that it took 2–10 years from project conception to finishing construction.

II.C. Water Pollution Background

This article emphasizes two measures of water quality—the dissolved oxygen saturation of water, and whether waters are fishable—though also reports results for other measures. We focus on dissolved oxygen saturation because it is among the most common omnibus measures of water quality in research, it responds to a wide variety of pollutants, and it is a continuous (rather than binary) measure of pollution, which alleviates concerns about failing to measure inframarginal changes in water quality. Most aquatic life requires dissolved oxygen to survive. Water can absorb dissolved oxygen from the air but loses dissolved oxygen when microorganisms consume oxygen to decompose pollution. Dissolved oxygen levels move inversely with temperature. Dissolved oxygen saturation represents the dissolved oxygen level divided by the maximum oxygen level expected given the water temperature, so implicitly adjusts for water temperature. Actual dissolved oxygen saturation is bounded below at zero (describing water with no oxygen) but is not bounded above. Dissolved oxygen deficits are defined as 100 minus dissolved oxygen saturation.

We focus also on the fishable standard because making water safe for fishing is a major goal of the Clean Water Act, and because recreational fishing is believed to be a main reason why people value water quality. We use a definition of “fishable” developed by William Vaughan for Resources for the Future (RFF). This definition distills several published water quality criteria and state water quality standards from between 1966 and 1979. It is also a widely used interpretation of “fishable.” In this definition, water is “fishable” if pollution is below a threshold, based on four measures: biochemical oxygen demand (BOD), dissolved oxygen saturation, fecal coliforms, and total suspended solids (TSS). To implement these definitions in the data, we pool data from these pollutants and define a dummy for whether a raw pollution reading exceeds the relevant standard.⁷

We also report estimates for whether waters are swimmable, and we report separate results for the other pollutants that are part of the “fishable” and “swimmable” definitions—BOD, fecal coliforms, and TSS. These pollutants merit interest in their own right because BOD, fecal coliforms, and TSS are a majority of the five “conventional pollutants” the Clean Water Act targeted. The other “conventional” pollutants are pH, which we analyze in Online Appendix Table IV, and oil and grease, a pollutant for which we have little data. We define all pollutants so that lower levels of the pollutant represent cleaner water (so we report the share of waters that are “not fishable” or “not swimmable,” and we report dissolved oxygen deficits).

Describing these other pollutants may help interpret results. BOD measures the amount of oxygen consumed by decomposing organic matter. Fecal coliforms proxy for the presence of pathogenic bacteria, viruses, and protozoa like E. coli that cause human illness. Pathogens including fecal coliforms are the most common reason water quality violates state standards today (USEPA 2016). TSS measures the quantity of solids in water that is trapped by a filter.⁸ Municipal sources in the early 1980s were estimated to account for about 20% of national BOD emissions and less than 1% of national TSS emissions (Gianessi and Peskin 1981), though municipal sources may account for a larger share of emissions in urban areas. Most TSS comes from agriculture and urban runoff.

We also report a few results for three additional groups of pollutants: industrial pollutants like lead, mercury, and phenols; nutrients like nitrogen and phosphorus; and other general water quality measures like temperature. We use a standardized criterion, described in Online Appendix B.3, to choose pollutants for this appendix table.

One important question is how far these pollutants travel downstream. We focus on a distance of 25 miles for several reasons. First, the only engineering study we found on this question (USEPA 2001) limited its analysis to 25 miles downstream of point sources for BOD. They chose this distance to reflect 15 watershed-specific studies designed to remedy pollution problems. Second, an interview with a wastewater regulation specialist at the Iowa Department of Natural Resources suggested that effects of treatment plants on dissolved oxygen would be concentrated within 20 miles downriver. Third, estimated effects of grants on whether rivers are fishable out to 100 miles downstream of a treatment plant only show effects within 25 miles (Online Appendix Table VI).

III. Data

We use eight types of data; Online Appendix B provides additional details.

III.A. Spatial Data on Rivers and Lakes

We use data from the National Hydrography Dataset Plus, Version 2.1 (NHD), an electronic atlas mapping all U.S. surface waters. NHD organizes the United States into approximately 200 river basins, 2,000 watersheds, 70,000 named rivers, 3.5 million stream and river miles, and 70 million river nodes. A river in these data consists of a set of river nodes (i.e., points) connected by straight lines. NHD forms a network describing the flow direction of each river or stream segment and helps us follow water pollution upstream or downstream. Figure I, Panel A shows U.S. streams, rivers, and lakes, colored by their distance from the ocean, Great Lakes, or other terminus. (See details in Online Appendix B.2.)

Figure I

Open in new tabDownload slide

National Maps of Water Pollution Data

In Panel A, rivers are colored by Stream Level from the National Hydrography Dataset. Streams that flow into oceans, the Great Lakes, Canada, or Mexico are the darkest. Streams that flow into these are lighter; streams that flow into these are still lighter, and so on. Panel B includes wastewater treatment plants used in the analysis (continental United States, within 1 km of a river, etc.). Panel C shows monitoring sites appearing in the years 1962–2001.

III.B. Municipal Water Pollution Sources

We use data on U.S. municipal water pollution treatment plants from the EPA’s Clean Watershed Needs Survey (CWNS). We use latitude and longitude data from the first available year for a plant (CWNS reports this beginning in 1984), and grant identifying codes for all available years. We limit the analysis to plants that report nonzero population served.

III.C. Clean Water Act Grants

We filed two Freedom of Information Act requests to obtain details on each of the 35,000 Clean Water Act grants the federal government gave to these plants. These records come from the EPA’s Grants Information and Control System (GICS). We restrict the analysis to grants with nonmissing award date, grant amount, and total project cost (including both federal and local capital expenditures). The data also report the name of the overseeing government authority (city, county, state, or special district), a grant identifier code, and the name of the recipient treatment plant. The data also include grants in the years 1957–1971 given under predecessor laws to the Clean Water Act. For simplicity, the analysis counts multiple grants to a treatment plant in a calendar year as a single grant. (See details in Online Appendix B.4.)

III.D. Ambient Levels of Water Pollution

We use water pollution readings from three federal data repositories: Storet Legacy, Modern Storet, and the National Water Information System (NWIS).⁹ Storet Legacy focuses on the earlier part of our period, and the full raw data include 18,000 data files and 200 million pollution readings. Modern Storet is similar to Storet Legacy but covers more recent years. The Storet repositories have data from many local organizations. USGS national and state offices collect a large share of NWIS readings. Online Appendix B.3 describes details and steps taken to clean these data, including limiting to rivers, streams, or lakes; restricting to comparable measurement methods; winsorizing at the 99th percentile; excluding readings specific to hurricanes and other nonroutine events; and more.

Online Appendix Table I provides basic descriptive statistics. The analysis sample includes 11 million observations on the four main pollutants and 38 million observations on the additional pollutants discussed in Online Appendix Table IV. The analysis sample covers 180,000 monitoring sites; an additional 60,000 monitoring sites record data on the other pollutants discussed in Online Appendix Table IV. Levels of BOD, fecal coliforms, and dissolved oxygen deficits are much lower in the United States than in India or China (Greenstone and Hanna 2014). Among the four main pollutants, about half the data describe dissolved oxygen. Almost half the data come from monitoring sites that report readings in at least three of the four decades we analyze.

No sampling design explains why certain areas and years were monitored more than others. In some cases, hydrologists purposefully designed representative samples of U.S. waters. At least three such networks are in these data: the Hydrologic Benchmark Network, the National Stream Quality Accounting Network, and the National Water-Quality Assessment Program (HBN, NASQAN, and NAWQA), which this article discusses later. In other cases, sampling locations and frequency were chosen by local governments or nongovernmental organizations. Cities and some states like Massachusetts have denser monitoring networks, while other areas like Texas have less dense networks (Figure I, Panel C).

III.E. Census Tract Data

We use the Geolytics Neighborhood Change Database (NCDB), which Geolytics built from the 1970, 1980, 1990, and 2000 Censuses of Population and Housing. The 1970 census only included metro areas in tracts, so these tract-level data for 1970 are restricted to metro areas, and so much of our analysis is as well.

We use these census data because they have national coverage and because transaction-level records from county assessor offices, such as those aggregated by Dataquick or CoreLogic, generally do not extend back to the 1970s. Online Appendix B.5 provides further details, including a discussion of data quality.

III.F. Recreational Travel Distances

We seek to determine a distance around a river that covers most individuals who travel to participate in recreation at this river. We obtain estimates of this distance from the Nationwide Personal Transportation Survey (NPTS) for 1983, 1990, and 1995. This survey is the only source we know that provides a large nationally representative sample of recreational activities and travel distances over the period we study.¹⁰ The survey picks a day and has respondents list all trips, their purposes, and the driving distances in miles. We limit trip purposes to “vacation” or “other social or recreation.” Averaged across the three survey years, the 95th percentile of one-way distance from home to recreational destinations is about 34 miles. Of course, these data represent all recreational trips and do not distinguish whether water-based recreation trips require different travel distances.

This is the distance traveled along roads, but the radius we use to calculate the distance of homes from rivers represents the shortest direct path along the ground (“great circle distance”). We are aware of two comparisons between great circle and road distances. First, the 2009 National Household Travel Survey (USDOT 2009, successor to the NPTS) reports both the road and great circle distance between a person’s home and the person’s workplace. The mean ratio of the road distance to the great circle distance is 1.4. Second, a recent study compared driving distance versus great circle distances for travel from a representative sample of 70,000 locations in the United States to the nearest community hospital, and the average ratio was also 1.4 (Boscoe, Henry, and Zdeb 2012).¹¹ So we estimate that the great circle distance between homes and rivers which covers 95% of recreational trips is 25 miles |$\left(\approx \frac{33.7}{1.4}\right)$|⁠.

III.G. Municipal Financial Records

To examine the pass-through of federal Clean Water Act grants to municipal spending on wastewater treatment, we use data from the 1970–2001 Annual Survey of State and Local Government Finances and the Census of Governments. These data report annual capital and total expenditures for sewerage (a category including wastewater treatment), separately for each local government.¹² The final sample includes 198 cities; in addition to describing these data in more detail, Online Appendix B.6 discusses the main sample restrictions, including requiring a balanced panel and accurate links to the grants data. Given this sample size, we report a set of estimates that weight by the inverse propensity score to provide estimates more representative of all cities. For use as a control variable in some specifications, we obtain population data for most of these cities from the 1970–2000 decennial censuses, then linearly interpolate between years.

III.H. Other Environmental Data

One sensitivity analysis controls for nearby industrial sources of water pollution. We are not aware of any complete data on industrial water pollution sources around 1972, so we use two distinct controls as imperfect proxies. The first is a list of the manufacturing plants that used large amounts of water in 1972. We obtain these data from the confidential 1973 Survey of Water Use in Manufacturing (SWUM) microdata, accessed through a Census Research Data Center. The second control is a count of the cumulative number of plants in a county holding industrial effluent (NPDES) permits. We filed a Freedom of Information Act request to obtain a historic copy of the EPA database that keeps records of industrial pollution sources—the Permit Compliance System, now called the Integrated Compliance Information System. Online Appendix B.7 describes more information on these sources, along with additional data on weather and nonattainment designations. Finally, Online Appendix B.8 describes data used to consider heterogeneity across different groups of grants by several dimensions: grant size, baseline abatement technologies, baseline pollution, Clean Water Act state decentralization, prevalence of local outdoor fishing and swimming, local environmental views, declining older urban areas (Glaeser and Gyourko 2005), and high-amenity areas (Albouy 2016).

1. Spatial Links. We construct four types of links between data sets. The first involves linking each pollution monitoring site and treatment plant to the associated river or lake. The second involves measuring distances along rivers between treatment plants and pollution monitoring sites. The third involves measuring areas of census tracts around rivers. The fourth involves linking grants to individual plants in the CWNS. Online Appendix C provides details of each step.

IV. Econometric and Economic Models

IV.A. Econometrics: Water Pollution Trends

Except where otherwise noted, all regressions in the article are clustered by watershed. Online Appendix Tables III, VI, and VIII also report results from two-way clustering by watershed and year. A watershed is defined by the USGS as an area of land in which all water within it drains to one point. Where relevant, watersheds or counties are defined by the treatment plant’s location.

IV.B. Econometrics: Effects of Grants on Water Pollution

This section discusses estimates of how grants affect downstream water pollution, which is the article’s second main research question. It then assesses how grants affect municipal spending on wastewater treatment capital. Online Appendix D discusses evidence on how water pollution changes as rivers pass treatment plants, which tests the hypothesis that the data capture an important feature of the world.

1. Effects of Clean Water Act Grants on Water Pollution

Equation (3) includes several important sets of controls. The matrix |$X_{pdy}^{^{\,\,\prime }}$| includes temperature and precipitation controls. The plant × downstream fixed effects η_pd allow both upstream and downstream waters for each treatment plant to have different mean levels of water pollution. These fixed effects control for time-invariant sources of pollution like factories and farms, which may be only upstream or only downstream of a plant. The plant × year fixed effects η_py allow for water pollution to differ near each treatment plant in each year, and they control for forces like the growth of local industries, other environmental regulations, and changes in population density which affect both upstream and downstream pollution. The downstream × basin × year fixed effects η_dwy allow upstream and downstream water quality separately to differ by year in ways that are common to all plants in a river basin. These fixed effects address the possibility that other point source pollutants and regulations are located near waste-water treatment plants and had water quality trends related to the municipal grants.

Equation (3) focuses on the effect of the number of grants a plant has received, rather than the dollar value of these grants, for several reasons. (Online Appendix Table VI reports similar effects of grant dollars.) First, it may be easier to think in discrete terms about the effect of a grant, rather than the effect of an arbitrary amount of money. Second, estimating these regressions in simple discrete terms makes the regression tables more easily comparable with event study graphs. Third, larger grants tend to go to more populated areas and larger rivers. Because it takes larger investment to achieve a change in pollution concentration for a more populated area and larger river, it is ambiguous whether larger grants should have larger effects on pollution concentrations. Fourth, the distribution of cumulative grant amounts is skewed and has many zeros. Focusing on the number of grants rather than grant dollars avoids issues involved in log transformations (or other approaches) in the presence of many zeros.

A few other details are worth noting. Because the dependent variable is an average over different numbers of underlying pollution readings, in all regressions where each observation is a plant-downstream-year tuple, we use generalized least squares (GLS) weighted by the number of raw underlying pollution readings.¹³ To maximize comparability between the treatment plant location and monitoring sites, we restrict pollution data to monitoring sites located on the same river as the treatment plant. Finally, estimates are limited to plants within one kilometer of a river node. Online Appendix Table VI shows results with some of these assumptions relaxed.

2. Pass-through of Clean Water Act Grants to Municipal Expenditure

How does |${\$}$|1 of Clean Water Act grants affect municipal spending on wastewater treatment? Grants could have complete pass-through, so a federal grant of |${\$}$|1 increases municipal spending on wastewater treatment by |${\$}$|1. Grants could also have incomplete pass-through (crowding out municipal expenditure) or more than complete pass-through (crowding in).

We study this question primarily because it can increase the accuracy of cost-effectiveness and cost-benefit analyses. If, for example, an additional |${\$}$|1 of federal grant funds lead cities to spend less than |${\$}$|1 on wastewater treatment, then the spending due to grants is less than our cost data imply. The question of how federal grants affect municipal spending is also important in the fiscal federalism literature (Oates 1999; Lutz 2010). Finally, this analysis provides some evidence on the quality of the grants data, because the grants data come from a completely different source than the municipal expenditure data.

The definitions of these variables are important. Municipal expenditures E_cy include both expenditures funded by federal grants and those funded by other sources of revenue. As mentioned in Section II.B, most grants require cities to pay 25% of the capital cost, though a small share require other copayments. We therefore report two sets of regressions—one where the variable D_cy includes only federal grant funds, and another where the variable D_cy includes both federal grant funds and the required municipal capital contribution. We also report specifications that weight by the inverse propensity score for inclusion in the balanced panel of cities.

IV.C. Demand for Water Quality

1. Hedonic Model

A few definitions and a graph convey essential features of the hedonic model. A house i is described by a vector of its J different characteristics, (z₁, …, z_J). The home’s price is P_i = P(z₁, …, z_J). The marginal implicit price of attribute j is the marginal change in home price due to a marginal increase in attribute j, all else constant: |$P_{z_{j}}\equiv \frac{\partial P}{\partial z_{j}}$|⁠. The key feature of this hedonic price schedule P(·) is that it reflects the equilibrium of firms that supply housing and consumers that demand housing. We assume that housing markets are competitive and that each consumer rents one house.

Online Appendix Figure VII illustrates. The curve θ₁ describes the bid function of one type of consumer. The bid function is the consumer’s indifference curve in the trade-off between the price of a home and the amount of attribute j embodied in the home. The curve θ₂ describes the bid function for another type of consumer. The curve φ₁ describes the offer function of a firm, and φ₂ of another firm. The offer function is the firm’s isoprofit curve in the trade-off between home price and attribute j.

The hedonic price schedule provides information about willingness to pay for amenity j because it reflects the points of tangency between consumer bid curves and firm offer curves. This implies that the marginal implicit price of an amenity at a given point on the hedonic price schedule equals the marginal willingness to pay of the consumer who locates on that point of the hedonic price schedule.

2. Econometrics: Demand for Water Quality

To analyze how Clean Water Act grants affected home values, we use a difference-in-differences estimate comparing the change in the log mean value of homes within a 0.25-, 1-, or 25-mile radius in any direction of the downstream river, before versus after the plant receives a grant, and between plants receiving grants in early versus late years.

Because water pollution flows in a known direction, areas upstream of a treatment plant provide a natural counterfactual for areas downstream of a plant. For this reason, our preferred methodology in Section IV.B to assess how Clean Water Act grants affect water pollution uses a triple-difference estimator comparing upstream and downstream areas. But because residents who live upstream of treatment plants can benefit from clean water downstream of treatment plants (e.g., by traveling for recreation), upstream homes could benefit from grants. Hence our preferred housing estimates come from difference-in-differences regressions analyzing homes within a 25-mile radius of river segments that are downstream of treatment plants. We report both the double-difference and triple-difference estimators for both outcomes, and obtain qualitatively similar conclusions.

A few points are worth noting. First, we limit regression estimates to the set of tracts reporting home values in all four years 1970, 1980, 1990, and 2000. When we fit the change in home values, we do so both for only the balanced panel of tract-years reporting home values, and for all tract-years. Second, because the difference-in-differences specification used for home values does not use upstream areas as a counterfactual, it involves the stronger identifying assumption that areas with more and fewer grants would have had similar home price trends in the absence of the grants. In part for this reason, we focus on specifications including basin × year fixed effects and the interaction of baseline characteristics with year fixed effects. Estimates without the basin × year controls are more positive but also more sensitive to specification, which is one indication that the specification of equation (6) provides sharper identification. Fourth, to obtain regression estimates for the average housing unit and provide an efficient response to heteroskedasticity, we include GLS weights proportional to the number of total housing units in the plant-year observation and to the sampling probability.¹⁷

V. Water Pollution Trends

V.A. Main Results

We find large declines in most pollutants that the Clean Water Act targeted. Dissolved oxygen deficits and the share of waters that are not fishable both decreased almost every year between 1962 and 1990 (Figure II). After 1990, the trends approach zero. Year-by-year trends for the other pollutants in the main analysis—the share of waters that are not swimmable, BOD, fecal coliforms, and TSS—show similar patterns (Online Appendix Figure III).

Figure II

Open in new tabDownload slide

Water Pollution Trends, 1962–2001

Graphs show year fixed effects plus a constant from regressions that also control for monitoring site fixed effects, a day-of-year cubic polynomial, and an hour-of-day cubic polynomial, corresponding to equation (1) from the text. Connected dots show yearly values, dashed lines show 95% confidence interval, and 1962 is the reference category. Standard errors are clustered by watershed.

The graphs show no obvious evidence of a mean shift or trend break in water pollution around 1972. This tells us little about the Clean Water Act’s effects, however, since its investments may take time to affect water pollution, expanded during the 1970s, and may be effective even if not obvious from a national time series. These graphs also suggest that existing evaluations of the Clean Water Act, which typically consist of national trend reports based on data from after 1972, may reflect forces other than the Clean Water Act. Using a national time series to evaluate the Clean Water Act could imply that it has been counterproductive, since the rate of decrease in pollution slowed after 1972.

Regressions with linear trend and trend break specifications underscore these findings, subject to the caveats mentioned earlier about the linear approximations and the long post period. The share of waters that are not fishable fell on average by about half a percentage point per year, and the share that are not swimmable fell at a similar rate (Table I, Panel A). In total over the period 1972–2001, the share of waters that are not fishable and the share not swimmable fell by 11 to 12 percentage points. Each of the four pollutants which are part of these fishable and swimmable definitions declined rapidly during this period. Fecal coliforms had the fastest rate of decrease, at 2.5% a year. BOD, dissolved oxygen deficits, and total suspended solids all declined at 1% to 2% a year.

These full data show more rapid declines before 1972 than after it. Independent evidence is generally consistent with this idea. Engineering calculations in USEPA (2000c) suggest that the efficiency with which treatment plants removed pollution grew faster in the 1960s than in the 1980s or 1990s. Hines (1967) describes state and local control of water pollution in the 1960s, which typically included legislation designating regulated waters and water quality standards, a state pollution control board, and enforcement powers against polluters including fines and incarceration. Data on industrial water pollution in the 1960s is less detailed, though manufacturing water intake (which is highly correlated with pollution emissions) was flat between 1964 and 1973 due to increasing internal recycling of water (Becker 2016). Moreover, the share of industrial water discharge that was treated by some abatement technology grew substantially in the 1960s (U.S. Census Bureau 1971). We interpret pre-1972 trends cautiously, however, because far fewer monitoring sites recorded data before the 1970s (Online Appendix Table I) and because the higher-quality monitoring networks (NAWQA, NASQAN, and HBN) focused their data collection after 1972.

It is interesting to consider possible explanations for these slowing trends. One involves declining returns to abatement of pollution from “point sources.” At the same time, much oxygen-demanding pollution comes from agriculture and other “nonpoint” sources, and those sources have remained largely unregulated. Another is that “fishable” and “swimmable” are limited between 0 and 1, and dissolved oxygen saturation does not much exceed 100%. This explanation is less relevant for the slowing trends in continuous variables like BOD, fecal coliforms, or TSS.

We estimate many sensitivity analyses, including restricting to high-quality subsamples of the data, adding important controls, weighting by population, and many others. Most of these alternative approaches have similar sign, magnitude, and precision as the main results. Online Appendix Table III shows these results and Online Appendix E.1 explains each.

V.B. Other Water Quality Measures

We also discuss trends in three other groups of water quality measures: industrial pollutants, nutrients, and general measures of water quality (Online Appendix Table IV).¹⁸ All three industrial pollutants have declined rapidly. Lead’s decrease of about 10% a year may be related to air pollution regulations, such as prohibiting leaded gasoline. The decline in mercury is noteworthy given the recent controversy of the Mercury and Air Toxics Standards (MATS) policy that would regulate mercury from coal-fired power plants. Some nutrients like ammonia and phosphorus are declining, while others like nitrates are unchanged. Nutrients were not targeted in the original Clean Water Act but are a focus of current regulation. Temperature is increasing by about 1°F per 40 years, which is consistent with effects from climate change. Electricity-generating units and other sources do contribute to thermal pollution in rivers, but increasing temperature is an outlier from decreasing trends in most other water pollutants.

pH increased by 0.007 pH units a year, meaning that waters became more basic (less acidic). Rainwater monitors that are not in our data record increases of similar magnitude in rainwater pH over this period, and attribute it to declines in atmospheric sulfur air pollution (USEPA 2007). Hence decreases in acidic sulfur air pollution may have contributed to decreases in acidic water pollution.

VI. Clean Water Act Grants and Water Pollution

VI.A. Effects of Clean Water Act Grants on Pollution

Table II shows that these grants cause large and statistically significant decreases in pollution. Each grant decreases dissolved oxygen deficits by 0.7 percentage points, and decreases the probability that downstream waters are not fishable by 0.7 percentage points. The other pollutants decrease as well—BOD falls by about 2.4%, fecal coliforms fall by 3.6%, and the probability that downstream waters are not swimmable by about half a percentage point. The point estimate implies that each grant decreases TSS by 1%, though this is imprecise.

Event study graphs corresponding to equation (4) support these results. In years before a grant, the coefficients are statistically indistinguishable from zero, have modest magnitude, and have no clear trend (Figure III). This implies that pollution levels in upstream and downstream waters had similar trends before grants were received. In the years after a grant, downstream waters have 1–2% lower dissolved oxygen deficits, and become 1–2% less likely to violate fishing standards. These effects grow in magnitude over the first 10 years, are statistically significant in this period, and remain negative for about 30 years after a grant. The gradual effect of the grants is unsurprising since, as mentioned earlier, the EPA estimates that it took 2 to 10 years after a grant was received for construction to finish. The 30-year duration of these benefits is also consistent with, though on the lower end of, engineering predictions. Two studies report that concrete structures of treatment plants are expected to have a useful life of 50 years, but mechanical and electrical components have a useful life of 15–25 years (USEPA 2002, 11; American Society of Civil Engineers 2011, 15). Event study graphs for other pollutants are consistent with these results, but are less precise (Online Appendix Figure IV). Online Appendix Figure V shows the effect of a grant by distance downstream from a treatment plant; few data are available to estimate effects separately for each five-mile bin along the river, and estimates are correspondingly less precise.

Figure III

Open in new tabDownload slide

Effects of Clean Water Act Grants on Water Pollution: Event Study Graphs

Graphs show coefficients on downstream times year-since-grant indicators from regressions which correspond to the specification of Table II. These regressions are described in equation (4) from the text. Data cover the years 1962–2001. Connected dots show yearly values, dashed lines show 95% confidence interval. Standard errors are clustered by watershed.

Online Appendix Table VI shows a variety of sensitivity analyses, and Online Appendix E.2 discusses each. They give similar qualitative conclusions as the main results, though exact point estimates vary.

VI.B. Grants’ Effects on Water Pollution: Cost-Effectiveness

We now turn to estimate the cost-effectiveness of these grants. The cost-effectiveness is defined as the annual public expenditure required to decrease dissolved oxygen deficits in a river-mile by 10 percentage points or to make a river-mile fishable. These calculations use our regression estimates and the cost data.

Even without the hedonic estimates of the next section, one can combine cost-effectiveness numbers with estimates from other studies of the value of clean waters to obtain a cost-benefit analysis of these grants. Moreover, we are not aware of any existing ex post estimates of the cost required to make a river-mile fishable or to decrease dissolved oxygen deficits.

Table III presents estimates of cost-effectiveness. The simplest specification of column (1), which includes rivers with water quality data, implies that it cost |${\$}$|0.67 million a year to increase dissolved oxygen saturation in a river-mile by 10%; the broadest specification of column (3), which assumes every treatment plant has 25 miles of downstream waters affected, implies that it cost |${\$}$|0.53 million a year. The annual cost to make a river-mile fishable ranges from |${\$}$|1.5 to |${\$}$|1.9 million.¹⁹

A few notes are important for interpreting these statistics. First, this is the average cost to supply water quality via Clean Water Act grants; the marginal cost, or the cost for a specific river, may differ. Second, measuring cost-effectiveness is insufficient to reach conclusions about social welfare; Section VII discusses peoples’ value for these changes. Third, if some grant expenditures were lost to rents (e.g., corruption), then those expenditures represent transfers and not true economic costs. The EPA did audit grants to minimize malfeasance. In the presence of such rents, this analysis could be interpreted as a cost-effectiveness analysis from the government’s perspective.

Online Appendix E.2 investigates heterogeneity in grants’ effects on water pollution and cost-effectiveness. Overall, this evidence does not suggest dramatic heterogeneity in cost-effectiveness. Compared to the mean grant, grants to declining urban areas are significantly less cost-effective, whereas grants to the generally rural counties where many people go fishing or swimming are significantly more effective. Most others are statistically indistinguishable from the mean grant, though there is some moderate (if statistically insignificant) heterogeneity in point estimates.

VI.C. Pass-through of Clean Water Act Grants to Municipal Expenditure

Table IV reports estimates corresponding to equation (5). In Panel A, the main explanatory variable excludes required municipal contributions, while Panel B includes them. Column (1) reports a basic difference-in-differences regression with nominal dollars. Column (2) uses real dollars. A city may spend a grant in years after it is received, so real pass-through may be lower than nominal pass-through. Column (3) adds river basin × year fixed effects. Column (4) reweights estimates using the inverse of the estimated propensity score for inclusion in the balanced panel of cities.

The estimates in Table IV are generally consistent with near complete pass-through, that is, little or no crowding out or in beyond the required municipal capital copayment. Panel A estimates pass-through modestly above 1 since it excludes the required municipal copayment. Panel B includes the local copayment, and finds pass-through rates of 0.84 to 0.93 in real terms or 1.09 in nominal terms. These estimates are within a standard deviation of one, so fail to reject the hypothesis that the municipal wastewater investment exactly equals the cost listed in the grant project data.²⁰

We emphasize a few caveats in interpreting Table IV. First, the analysis is based on only 198 cities. The inverse propensity score reweighted estimates are designed to reflect the entire population of U.S. cities. Second, this city-level difference-in-differences estimate cannot use the upstream-downstream comparison for identification. Third, this analysis is different from the question of what municipal spending (and pollution and home values) would be in a world without the Clean Water Act. Our estimates are consistent with no crowding out for an individual grant, but the existence of the Clean Water Act may decrease aggregate municipal investment in wastewater treatment. Online Appendix Figure VI shows national trends in federal versus state and local spending on wastewater treatment capital over 1960–1983.²¹ State and local spending on wastewater treatment capital declined steadily from a total of |${\$}$|43 billion in 1963 to |${\$}$|22 billion in 1971 and then to |${\$}$|7 billion annually by the late 1970s. Notably, almost half of this decline in state and local wastewater treatment capital spending occurred before the Clean Water Act. Federal spending grew to between |${\$}$|10 and |${\$}$|20 billion a year in the late 1970s.

Other sources note that these time series trends are consistent with aggregate crowding out (Jondrow and Levy 1984; CBO 1985). Identification from a national time series is difficult, since other national shocks like the 1973–1975 and early 1980s recessions, high inflation and interest rates, and the OPEC crisis make the 1960s a poor counterfactual for the 1970s and 1980s.

Our interpretation is that once the Clean Water Act began, cities became less likely to spend municipal funds on wastewater treatment capital. In this sense, the existence of the Clean Water Act did crowd out aggregate municipal investment in wastewater treatment. But municipal investments that occurred were closely connected to grants, and point estimates imply that the grant costs in our data accurately represent the actual change in spending. Online Appendix E.2 discusses how cost-effectiveness numbers change with alternative estimates of crowding out.²²

These pass-through estimates also speak to the broader “flypaper” literature in public finance, so named to reflect its finding that federal government spending “sticks where it hits.” Researchers have estimated the pass-through of federal grants to local expenditure in education, social assistance, and other public services. A review of 10 U.S. studies found pass-through estimates between 0.25 and 1.06 (Hines and Thaler 1995). Non-U.S. studies and more recent U.S. estimates find an even wider range (Gamkhar and Shah 2007). One general conclusion from this literature is that the effect of federal grants on local government expenditure substantially exceeds the effect of local income changes on local government expenditure (the latter is typically around 0.10). This literature also finds that federal grants that require local matching funds and specify the grants’ purpose, both characteristics of the Clean Water Act grants, tend to have higher pass-through rates. Our findings are consistent with these general conclusions.

VII. Demand for Water Quality

VII.A. Main Results

Table V analyzes how Clean Water Act grants affect housing. Column (1) shows estimates for homes within a quarter mile of downstream waters. Column (2) adds controls for dwelling characteristics, and for baseline covariates interacted with year fixed effects. Column (3) include all homes within 1 mile, and column (4) includes homes within 25 miles.

Panel A reports estimates of how grants affect log mean home values. The positive coefficients in the richer specifications of columns (2) through (4) are consistent with increases in home values, though most are statistically insignificant. Column (4) implies that each grant increases mean home values within 25 miles of affected waters by 0.024 percentage point. The 0.25- or 1.0-mile estimates are slightly larger, which is consistent with the idea that residents nearer to the river benefit more from water quality. Panel B analyzes how grants affect log mean rental values. These estimates are even less positive than the estimates for housing. The estimate in column (4), including homes within a 25-mile radius of downstream rivers, is small and statistically insignificant but actually negative.

Panels A and B reflect the classic hedonic model, with fixed housing stock. Panel C estimates the effect of grants on log housing units and Panel D on the log of the total value of the housing stock. They suggest similar conclusions as Panels A and B. Most of these estimates are small and actually negative. Two are marginally significant (Panel C, column (1)), though the precision and point estimate diminish with the controls of column (2).

Figure IV shows event study graphs, which suggest similar conclusions as these regressions. Panel A shows modest evidence that in the years after a plant receives a grant, the values of homes within 0.25 mile of the downstream river increase. The increases are small and statistically insignificant in most years. Panel B shows no evidence that homes within 25 miles of the downstream river increase after a treatment plant receives a grant.

Figure IV

Open in new tabDownload slide

Effects of Clean Water Act Grants on Log Mean Home Values: Event Study Graphs

Graphs show coefficients on year-since-grant indicators from regressions corresponding to the specification of Table V, columns (2) and (4). Connected dots show yearly values, dashed lines show 95% confidence interval. Standard errors are clustered by watershed. Panels A and B show different ranges of values on their y-axes. Data cover decennial census years 1970–2000.

We also report a range of sensitivity analyses, which are broadly in line with the main results. Estimates appear in Online Appendix Table VIII and discussion appears in Online Appendix E.3.

VII.B. Measured Benefits and Costs

We now compare the ratio of a grant’s effect on housing values (its “measured benefits”) to its costs. The change in the value of housing is estimated by combining the regression estimates of Table V with the baseline value of housing and rents from the census. Grant costs include local and federal capital expenditures plus operating and maintenance costs over the 30-year life span for which we estimate grants affect water pollution. We deflate operating and maintenance costs and rents at a rate of 7.85% (Peiser and Smith 1985).²³

Column (1) of Table VI includes only owned homes within a 1-mile radius of the downstream river segments; column (2) includes homes within a 25-mile radius; and column (3) adds rental units. Column (4) includes imputed home values for the nonmetro areas that were not in the 1970 or 1980 census.²⁴

Considering all owner-occupied homes within 25 miles of the river, the estimated ratio of the grants’ aggregate effects on home values to the grants’ costs is 0.26. Adding rental units in column (3) barely changes this estimate. The main regression sample includes only a balanced panel of tracts that appear in all four censuses between 1970 and 2000; imputing values for missing homes hardly changes the ratio in column (4). These confidence regions do not reject the hypothesis that the ratio of the change in home values to the grants’ costs is zero but do reject the hypothesis that the change in home values equals the grants’ costs.

Online Appendix Table VII investigates heterogeneity in measured benefits and costs; Online Appendix E.3 discusses the results. We find suggestive evidence that ratios of measured benefits to costs follow sensible patterns, though not all estimates are precise. None of these subsets of grants considered has a ratio of measured benefits to costs above one, though many of the confidence regions cannot reject a ratio of 1. The largest ratios of estimated benefits to costs are for areas where outdoor fishing or swimming is common (ratio of 0.53), for high-amenity urban areas (ratio of 0.40), and in the South (ratio of 0.84).

The map in Online Appendix Figure VIII shows heterogeneity in the ratio of measured benefits to costs across U.S. counties. This map assumes the same hedonic price function and reflects spatial heterogeneity in housing unit density.²⁵ The map shows that the ratio of measured benefits to costs is larger in more populated counties. The bottom decile of counties, for example, includes ratios of measured benefits to costs of below 0.01. The top decile of counties includes ratios between 0.31 and 0.41. Grants and population are both skewed, so large shares of both are in the top decile. Although a point estimate of 0.41 for the ratio of benefits to costs does not exceed 1, one should interpret this value in light of the discussion from the next subsection that it may be a lower bound on true benefits.

This predictable spatial variation in the net benefits of water quality variation suggests that allowing the stringency of regulation to vary over space may give it greater net benefits (Muller and Mendelsohn 2009; Fowlie and Muller forthcoming).

VII.C. Interpreting Hedonic Estimates

We now discuss six reasons the ratios of measured benefits to costs from the previous subsection may provide a lower bound on the true benefit/cost ratio. Online Appendix F discusses other reasons we believe have weaker support.

First, people might have incomplete information about changes in water pollution and their welfare implications. Research does find statistically significant but imperfect correlation between perceived local water pollution and objectively measured local water pollution (Faulkner et al. 2001; Jeon et al. 2011; Poor et al. 2001; Steinwender, Gundacker, and Wittmann 2008; Artell, Ahtiainen, and Pouta 2013). Incomplete information would be especially important if pollution abatement improves health. Misperception would be less important if most benefits of surface water quality accrue through recreation or aesthetics, since failing to perceive water pollution through any means would mean its effects on recreational demand are limited. Most recent cost-benefit analyses of the Clean Water Act estimate that a substantial share of benefits come from recreation and aesthetics channels (Lyon and Farrow 1995; Freeman 2000; USEPA 2000a). Cropper and Oates (1992) describe the Clean Water Act as the only major environmental regulation of the 1970s and 1980s that does not have health as its primary goal.

Second, due to “nonuse” or “existence” values, a person may value a clean river even if they never visit or live near that river. We recognize the potential importance of nonuse values for clean surface waters and the severe challenges in accurately measuring these values.²⁶ Other categories potentially not measured here include the value for commercial fisheries, industrial water supplies, lower treatment costs for drinking water, and safer drinking water.²⁷ Evidence on the existence and magnitude of the benefits from these other channels is limited, though as mentioned already, recreation and aesthetics are believed to account for a large majority of the benefits of clean surface waters.

Third, these grants could lead to increased city taxes, sewer fees, or other local costs that depress home values. Table VI separately lists three types of costs: federal expenditures on capital, local expenditures on capital, and operation and maintenance costs. The ultimate entity responsible for local capital costs and operation and maintenance costs is ambiguous because local governments may receive other payments from state or federal governments to help cover these costs. But if local governments ultimately pay these costs, they could depress home values.

A few pieces of evidence help evaluate the relevance of these issues. One is to estimate hedonic regressions excluding housing units in the same city as the wastewater treatment plant. This is potentially informative because increased taxes, sewer fees, or changes in other municipal expenditures are likely to be concentrated in the municipal authority managing the treatment plant, whereas the change in water quality is relevant for areas further downstream. Row 12 of Online Appendix Table VIII reports this specification and finds similar and if anything slightly less positive change in home values than the main results estimate, which is the opposite of what one would expect if city taxes, sewer fees, or other local costs depressed home values. Another test comes from the fact that the 1980–2000 gross rent data reported in the census include utilities costs. If sewer fees were particularly important, then one would expect rents to increase more than home values do; if anything, the estimates of Table V suggest the opposite. Finally, we can recalculate the ratios in Table VI considering only subsets of costs. The ratio of the change in housing values to federal capital costs in columns (2)–(4) of Table VI ranges from 0.8 to 0.9; the ratio of the change in housing values to the sum of federal capital costs and operating costs (but excluding local capital costs) in these columns is around 0.3. None of these ratios exceeds 1, though they are closer to 1 than are the values in Table VI.

Fourth, this analysis abstracts from general equilibrium changes. One possible channel is that wages change to reflect the improvement in amenities (Roback 1982). A second general equilibrium channel is that the hedonic price function may have shifted. In the presence of such general equilibrium changes, our estimates could be interpreted as a lower bound on willingness to pay (Banzhaf 2015).

Other possible general equilibrium channels describe reasons the effects of cleaning up an entire river system could differ from summing up the effects of site-specific cleanups. One such channel involves substitution—cleaning up part of a river in an area with many dirty rivers might have different value than cleaning up a river in an area with many clean rivers. Another possible channel involves ecology. The health of many aquatic species (so indirectly, the benefit people derive from a river) may depend nonlinearly on the area of clean water. Our approach focuses on the effects of cleaning up an individual site and is not as well suited to capture the potentially distinct effects of cleaning up entire river systems.

Fifth, the 25-mile radius is only designed to capture 95% of recreational trips. The last 5% of trips might account for disproportionate surplus because they represent people willing to travel great distances for recreation. Alternatively, the most distant travelers might be marginal. Our recreation data also represent all trips, and water-based recreation trips might require different travel distances.

Finally, we interpret our pass-through estimates cautiously because they reflect only 198 cities, do not use upstream waters as a comparison group, and reflect pass-through of marginal changes in investment, rather than the entire Clean Water Act. Online Appendix E.3 discusses interpretations of our housing estimates under alternative pass-through numbers.

VIII. Conclusions

This article assembles an array of new data to assess water pollution’s trends, causes, and welfare consequences. We find that by most measures, U.S. water pollution has declined since 1972, though some evidence suggests it may have declined at a faster rate before 1972. The share of waters that are fishable has grown by 12 percentage points since the Clean Water Act.

We study |${\$}$|650 billion in expenditure from 35,000 grants the federal government gave cities to improve wastewater treatment plants. Each grant significantly decreased pollution for 25 miles downstream, and these benefits last for around 30 years. We find weak evidence that local residents value these grants, though estimates of increases in housing values are generally smaller than costs of the grant projects.

Our estimated ratio of the change in housing costs to total grant costs may provide a lower bound on the true benefit/cost ratio of this grant program because we abstract from nonuse (“existence”) values, general equilibrium effects, potential changes in sewer fees, and the roughly 5% longest recreational trips. The point estimates imply that the benefits of the Clean Water Act’s municipal grants exceed their costs if these unmeasured components of willingness to pay are three or more times the components of willingness to pay that we measure. As mentioned in the introduction, other recent analyses estimate benefits of the Clean Water Act that are smaller than its costs, though these other estimates note that they may also provide a lower bound on benefits. For example, the USEPA’s (2000a,b) estimate of the benefit/cost ratio of the Clean Water Act is below 1, though the EPA’s preferred estimate of the benefit/cost ratio of the Clean Air Act is 42 (USEPA 1997).²⁸

It may be useful to highlight differences in how the Clean Air and Clean Water Acts answer four important questions about environmental regulation. These comparisons also highlight features of the Clean Water Act that are not widely recognized and could lead it to have lower net benefits than some other environmental regulation.

First is the choice of policy instrument. Market-based instruments are believed to be more cost-effective than alternatives. Parts of the Clean Air Act use cap-and-trade systems, but nearly none of the Clean Water Act does. The grants we study actually subsidize the adoption of pollution control equipment, which is a common policy that has undergone little empirical economic analysis.

A second question is scope. Cost-effective regulation equates marginal abatement costs across sources, which requires regulating all sources. The Clean Air Act covers essentially all major polluting sectors. The Clean Water Act, by contrast, mostly ignores nonpoint pollution sources like agriculture. Ignoring such a large source of pollution can make aggregate abatement more costly.

A third question involves substitution. Optimizing consumers should equate the marginal disutility of pollution to the marginal cost of protection from pollution. People breathe the air quality where they live, and relocating to another airshed or some other defenses against air pollution are costly (Deschenes, Greenstone, and Shapiro 2017). For water pollution, however, people can more easily substitute between nearby clean and dirty rivers for recreation.

A fourth question involves health. Air is typically unfiltered when it is inhaled, so air pollution is believed to have large mortality consequences that account for much of the benefits of air pollution regulation. Surface waters, by contrast, are typically filtered through a drinking water treatment plant before people drink them. Most analyses of recent U.S. water quality regulation count little direct benefit from improving human health (Lyon and Farrow 1995; Freeman 2000; USEPA 2000a; Olmstead 2010).²⁹

Finally, we note one similarity between air and water pollution that may be relevant to policy design. We find some evidence that the net benefits of Clean Water Act grants vary over space in tandem with population density and the popularity of water-based recreation. Related patterns have been found for air pollution, and suggest that allowing the stringency of pollution regulation to vary over space has potential to increase social welfare.

Footnotes

Supplementary Material

Data and code replicating tables and figures in this article can be found in Keiser and Shapiro (2018), in the Harvard Dataverse, doi:10.7910/DVN/2JRHN6.

References