Data.

To construct an empirical measure of a public good, it is key to note the appropriate spatial scale. To account for factors affecting water quality, we contend that the ideal size is watershed based. Drainage basin sizes vary and watersheds can be subdivided into smaller areas defined by the forks of main rivers. Thus, we rely on US Geological Survey (USGS) methodology. USGS divides the United States into successively smaller subunits identified by unique hydrologic unit codes (HUCs). USGS names these units with 2–14 digits based on decreasing size class (39). We use the eight-digit HUC (HUC8) designation as the definition of a watershed for two main reasons. First, drainage basins correspond to the natural boundary for surface water flow. Second, nonprofit groups targeting water quality generally operate at a local scale and the HUC8 scale corresponds to the smallest area a single group is likely to affect. There are 2,264 HUC8 watersheds in the United States, averaging 700 mi² in land cover size.

We focus on DO to measure water quality. We obtained DO measurements for rivers and streams from the National Water Information System (USGS) and Storet (US EPA) databases (available at the National Water Quality Monitoring Council’s Water Quality Portal; https://www.waterqualitydata.us/). The steps followed to construct the water quality sample from these data are described in SI Appendix. We use a standard formula to convert DO in milligrams per liter to DO saturation (in percentage), and calculate DOD as 100 – DO (in percent saturation). This process yields 2,276,913 measurements during our study period. Finally, we aggregate by calculating yearly averages of all measurements within each watershed.

We use proportion swimmable and fishable as additional measures of water quality. Swimmable and fishable indicators are often based on several pollutants, namely biochemical oxygen demand, fecal coliforms, and total suspended solids. Since there are fewer measurements for these pollutants than for DO, using them to define swimmable and fishable considerably reduces the size of the sample available for estimation. Hence, we construct indicator variables for these designations based on minimum DO thresholds: 4.99 mg/L for fishable and 6.47 mg/L for swimmable.^# We calculate the proportion of DO measurements in a watershed that meet each threshold and use these as dependent variables in place of DOD in our models.

We assume that watershed groups provide an appropriate context for assessing the effectiveness of environmental groups because geographical boundaries define a watershed rather than arbitrary political boundaries, and hence oversight by watershed groups takes place within the correct spatial scale to account for relevant factors affecting water quality. Our data on water-focused nonprofit groups come from several sources. An initial search in Guidestar, an organization that gathers information about nonprofits, yielded a comprehensive list of groups working on “Water Resource, Wetlands Conservation, and Management.” The information retrieved includes date of incorporation, location, type of group, and the Employer Identification Number (EIN), which is the federal tax identifier. This information does not allow us to further differentiate groups by more specific objectives. We cross-referenced and supplemented this information with lists from EPA and River Network, a national group assisting organizations whose mission is protecting water resources. These two datasets do not contain the EIN number, so we manually searched for and added any missing numbers to our final list. The EIN number links this list to a database from the US Internal Revenue Service with tax return data for each group, including yearly revenues and expenditures.

We estimate the effect of water groups conditional on factors that impact water quality at the watershed level. First, we account for state and federal enforcement of discharges covered under the CWA by including the total number of discharge permit violations in a watershed in each year. Information on violations and location of facilities are from the EPA’s Enforcement and Compliance History Online database.

We account for federal programs that aim to improve water quality by including total expenditures in the state where a watershed is located in each year under three key programs: the Conservation Reserve Program (CRP), the EPA 319 Grant Program, and the Environmental Quality Incentives Program (EQUIP). Data on CRP payments were obtained from the US Department of Agriculture Farm Service Agency (https://www.fsa.usda.gov/programs-and-services/conservation-programs/reports-and-statistics/conservation-reserve-program-statistics/index), data on payments made under the EPA 319 program are from the Grants Reporting and Tracking System (https://iaspub.epa.gov/apex/grts/f?p=109:9118), and information on payments made under EQUIP contracts was provided by the Environmental Working Group (https://www.ewg.org/).

Land use is another factor affecting water quality. Agriculture is a leading cause of impairment of surface waters (40). In urban areas, impervious surfaces contribute to increased pollutant loads and surface runoff of contaminants and sediment, as well as larger variances in stream flow and temperature (41). We account for these effects by using land cover maps from the Multi-Resolution Land Characterization Consortium to calculate proportions of urban and agricultural land in each watershed for each year.

Precipitation is another determinant of water quality. Small amounts of precipitation wash pollutants into water bodies through runoff, whereas relatively large amounts of precipitation accelerate dilution of pollutants. We include mean precipitation in a watershed and its square for each year using data from the PRISM Climate Group (42).

We condition on environmental and other political preferences of the residents of a watershed, as they may affect political outcomes or other factors that can affect water quality. To the extent that these preferences are expressed through voting, they can be accounted for by using election outcomes. We use the proportion of votes for Republican candidates in US Senate races using county-level data on election results from the CQ Press Voting and Elections Collection, interpolating for years in which there were no Senate races, and reweighting to the HUC8 level based on the proportions of counties contained in a watershed.

Finally, demographic characteristics affect water quality. We include population density, per capita income (in 2008 dollars), percentage of population with a high school degree, percentage of the population that are white, home ownership rate, and unemployment rate. This information is at the county level from the US Census and the Bureau of Economic Analysis. Population, income, and unemployment are available for every year in our study. Ethnicity, educational attainment, and home ownership are available for 1990, 2000, and 2010, so we interpolate for intracensus years. We reweight and aggregate these data to the watershed level.

We construct a panel dataset of these variables for 1,131 HUC8 watersheds in the 48 mainland states during 1995–2008. Our study period corresponds to years with reliable data on water groups. Summary statistics (SI Appendix, Table S1) suggest that water quality and water group activity, and indeed most other watershed characteristics, change gradually over the study period, reflecting slow-moving underlying processes.

Empirical Strategy.

An ideal research design would use data from an experiment in which water groups are randomly placed across watersheds. This, of course, is not feasible, and in practice the main challenge for measuring causal effects of water group activity on water quality is that water groups do not locate randomly across watersheds. Indeed, these groups may locate where water quality is relatively poor, which may lead to more donations and higher expenditures. As a result, watershed group activity and water quality may be jointly determined, and ordinary least-squares estimates may not be consistent. To address this concern, we follow a growing number of studies that combine matching and panel methods, particularly fixed-effects estimation (43⇓⇓⇓⇓⇓⇓–50).

We define treated watersheds as having at least one water-focused group during the study period, and control watersheds as those that do not have any groups during the entire study period. We preprocess the data to make treated and control watersheds observationally similar before the study period by matching on time-invariant or pretreatment observable characteristics that affect water quality: 1995 values for CWA violations, federal programs, precipitation, proportions of urban and rural land, proportion of vote for Republican candidates, per capita income, population density, high school graduation, ethnicity, unemployment, and home ownership rate. Additionally, we match within state because states are responsible for impairment listing of water bodies in the watershed. This accounts for state-level characteristics that may have an impact on water quality. [This accounts for the EPA requiring each state to submit a list of all threatened and impaired waters during even-numbered years (51) and provides information on which state is responsible for listing when water bodies cross state lines.] Finally, we match on water quality (DOD) in 1995. This helps mitigate the concern that water-focused groups may locate in watersheds with poorer water quality, since treated and control watersheds used in the estimation sample have similar DOD measures at the beginning of the study period. More details on the matching procedure and the balance of the estimation sample are presented in SI Appendix.

For the postmatching balanced sample, our regression models for mean DOD, proportion swimmable, and proportion fishable in watershed i in year t are as follows:DODit=α1Group Activityit−1+CWAit−1α2+Xitα3+δs+τt+εit,Swimmableit=β1Group Activityit−1+CWAit−1β2+Xitβ3+ϕs+θt+uit,Fishableit=γ1Group Activityit−1+CWAit−1γ2+Xitγ3+ψs+ωt+vit.

Group Activity_it is measured three ways: total number of groups, total donations, and total expenditures in watershed i and year t. Donations and expenditures are expressed in 2008 dollars. We include lagged effects of water group activity and CWA permit violations because watershed impairment can be a slow-moving process, and these factors likely have an impact with a lag. The matrix X_it contains contemporaneous values of the explanatory variables discussed above. The vectors τ_t, θ_t, and ω_t are year fixed effects. The vectors δ_s, ϕ_s, and ψ_s are fixed effects for the state responsible for impairment listing of water bodies in the watershed. This is the appropriate level to introduce fixed effects because water body management and impairment listing decisions are made annually at the state level. We avoid using watershed fixed effects because they remove the signal we wish to measure in the data (for example, see www.g-feed.com/2012/12/the-good-and-bad-of-fixed-effects.html) and thus eliminate most power of our explanatory variables. Fixed effects are preferred to random effects because the former allow us to control for unobserved state-level heterogeneity that is assumed to be correlated with included explanatory variables. Random effects, in contrast, are assumed to be independent of included regressors.

Alternative Specifications.

Our main models use fixed effects for the state responsible for impairment listing of water bodies. We consider two additional fixed-effects structures. First, we use state–year fixed effects to allow for more flexible time-variant state heterogeneity. Second, we use fixed effects for combinations of states included in transboundary watersheds to account for watersheds that extend across state lines and hence are subject to different jurisdictions. The results (SI Appendix, Tables S3 and S4) are consistent with those from our preferred specifications, except for the coefficient for number of groups, which is not statistically significant when using transboundary fixed effects. We also estimated the models without lagged regressors and with all regressors lagged. The results are consistent with those presented in Table 1. Finally, we estimated the models without the conditioning variables. If our approach for addressing nonrandom water group activity is effective, the water group variables should have the same sign as in the fully specified models. The results show that this is the case (SI Appendix, Table S5).

There is considerable spatial and temporal variation in water quality monitoring, and local governments or private organizations commonly choose the location and frequency of sampling. Some states have denser monitoring networks, waters near populated areas tend to be overrepresented, and monitors drop in and out over time (52).^||,** We conduct several sensitivity tests to ensure that our results are not driven by these characteristics of the DOD data. First, we use only monitors that have at least 25 measurements, as these may have higher-quality data. Second, we only use measurements from well-documented and high-quality monitors from the National Water Quality Assessment network.** Results (SI Appendix, Tables S6 and S7) are consistent with those from the full dataset. We also use summer rather than yearly DOD averages. Additionally, 10 states did not consistently file impairment reports during the study period (Utah, Vermont, Pennsylvania, New York, Nebraska, Mississippi, Minnesota, Connecticut, Colorado, and Washington). If this is due to unreliable monitoring, the data from these states may be relatively poor, so we run our models without these states. Results for these sensitivity checks (SI Appendix, Tables S8 and S9) are consistent with those of our preferred specifications.

Finally, we assess the robustness of our results to a different estimation approach. We define the outcome of interest as the percentage change in DOD between 1996 and 2008, and consider the same treatment (presence of at least one watershed group during the study period). Instead of matching followed by fixed-effects regression, we use matching (with four nearest neighbors) and then nonparametrically estimate the average treatment effect (ATE). This approach hinges on the conditional independence (unconfoundedness) assumption that, conditional on observed covariates, treatment is independent of the outcome (details are available in SI Appendix). We check for robustness to alternative numbers of nearest neighbors (1⇓–3, 5). We also assess sensitivity to omitting matching covariates and conduct a falsification test by estimating the ATE on a pretreatment outcome (change in DOD between 1992 and 1995). Results are presented in SI Appendix, Table S10.

The results support those from our main models. Water groups had a positive and significant impact on water quality, as the rate increase of DOD is smaller by 1.8 percentage points in watersheds where water groups were located. This effect is robust to the number of nearest neighbors used to construct the counterfactual matches. Additionally, the results are robust to the exclusion of covariates used for matching. Finally, there is no effect on the pretreatment outcome, which suggests there is no selection on unobserved factors (53).

Our paper gives proof of concept for these watershed groups providing the public good of improved water quality. Stronger results could be obtained with additional data: (i) an annual survey of watershed groups project regarding water quality, in other words, better expenditure information including where and how the money is spent, or (ii) randomizing grant funding to these groups, or (iii) better sampling of water quality (more temporally consistent and more spatially representative). These data do not exist, but their collection would further this research in the area of private provision of public goods by environmental groups.