Estimating biotic integrity to capture existence value of freshwater ecosystems

Significance

Estimating the value of changes in water quality requires the definition of biophysical features that link changes in ecosystems to changes in social systems. Those linking features must be interpretable to people and serve as effective ecological indicators. This work defines a linking feature that is appropriate for capturing existence values in a forthcoming national stated preference survey of Clean Water Act regulations. Further, we modeled and spatially predicted this feature to account for the dependence of survey respondents’ preference on baseline aquatic conditions near them. We outline steps to provide insights on the mechanisms that will aid in forecasting changes in the linking feature in responses to candidate policy options.

Abstract

The US Environmental Protection Agency (EPA) uses a water quality index (WQI) to estimate benefits of proposed Clean Water Act regulations. The WQI is relevant to human use value, such as recreation, but may not fully capture aspects of nonuse value, such as existence value. Here, we identify an index of biological integrity to supplement the WQI in a forthcoming national stated preference survey that seeks to capture existence value of streams and lakes more accurately within the conterminous United States (CONUS). We used literature and focus group research to evaluate aquatic indices regularly reported by the EPA’s National Aquatic Resource Surveys. We chose an index that quantifies loss in biodiversity as the observed-to-expected (O/E) ratio of taxonomic composition because focus group participants easily understood its meaning and the environmental changes that would result in incremental improvements. However, available datasets of this index do not provide the spatial coverage to account for how conditions near survey respondents affect their willingness to pay for its improvement. Therefore, we modeled and interpolated the values of this index from sampled sites to 1.1 million stream segments and 297,071 lakes across the CONUS to provide the required coverage. The models explained 13 to 36% of the variation in O/E scores and demonstrate how modeling can provide data at the required density for benefits estimation. We close by discussing future work to improve performance of the models and to link biological condition with water quality and habitat models that will allow us to forecast changes resulting from regulatory options.

The goal of the US Clean Water Act (CWA; 33 U.S.C. §1251 et seq.) is to maintain and restore the chemical, physical, and biological integrity of US waters. The US Environmental Protection Agency (EPA) often quantifies the outcome of CWA regulations on aquatic resources with a water quality index (WQI; [1]). The WQI is often used to convey the suitability of surface waters for a range of human uses, such as boating, fishing, and swimming, with a unidimensional indicator comprised of chemical, physical, and biological water quality parameters (2). Currently, the WQI is the primary metric used by EPA to monetize water quality changes in benefit–cost analyses based on stated preference (SP) studies (3). However, EPA’s reliance on the WQI to estimate the benefit of proposed policies fails to recognize a recent shift in how the Agency analyzes CWA regulations. Specifically, estimated benefits have expanded from primarily human health and recreational uses to include impacts on biological integrity or condition, such as the composition of species within an ecosystem (4), for which the WQI can be a poor indicator. Thus, contemporary benefit–cost analysis of CWA regulations requires a metric that captures changes in biological condition beyond those that affect use values.

Society’s value for environmental quality that is not driven by human uses is known as “existence value” (5, 6). Existence value is a component of nonuse value derived from the satisfaction that people get from stewardship of the environment, even if they will never use the resource in question (5, 7). Importantly, values held by nonusers of a resource can be small per household but potentially exceed use-based value when aggregated across many households (8). Thus, failure to account for this source of benefit may materially underestimate total benefits of CWA regulations. Further, some research objectives require estimating use and nonuse values in a separable way (9), something that cannot be done with a single indicator of environmental quality. The only way to capture existence value is through SP studies, but even the best-designed SP surveys can be cognitively challenging for respondents (10), and identifying indicators that avoid conflation or confusion between use and existence values is a major challenge. For example, using fish populations as an indication of ecosystem health is likely to solicit willingness-to-pay responses motivated by both recreational use and a sense of stewardship for the environment.

To capture the existence portion of total economic value more accurately, EPA is developing a national SP study that will present several attributes of hypothetical policy outcomes in a discrete choice experiment framework (11).^* The location and quantity of surface waters improved will vary across surveys to collect data on how distance from the resource and scope of the improvements affect willingness to pay. While economic theory implies that use values should decrease as distance from the resource increases, there are no such implications for existence value. The rate at which existence value decays with distance is thus an empirical question. Answering that question requires separable water quality attributes for use and existence values and is a primary motivation for this research. The 100-point WQI is benchmarked by waterbodies’ suitability for a range of human uses and will capture use values on the survey. A second indicator for biological condition that captures existence value should satisfy standard criteria for choice experiment attributes and be cognitively separable from the quality of recreational experiences to meet our research objectives. We describe our process for selecting a metric of biological condition by subjecting candidate metrics to focus group research and comparing them based on characteristics that Schultz et al. (12) identified as necessary for ecological metrics to have for valid benefits estimation (Selecting an Index of Aquatic Biological Condition).

An additional challenge in resource valuation is that people’s willingness to pay for a given improvement in water quality generally depends on baseline conditions near them (13). Capturing this dependence in our analysis of the SP survey responses across the conterminous United States (CONUS) will require water quality data for streams and lakes at a sufficient extent and density to characterize conditions near participants. However, existing observational datasets of biological condition collected by academic or government institutions are only available for a small subset of waterbodies or at an insufficient density. Therefore, we also report on modeling that will interpolate observed values of biological condition to unsampled streams and lakes and impart the coverage needed for a national analysis of CWA regulations. A national dataset of this kind will provide resource economists with a consistent measure of aquatic biological condition for estimating benefits of ecosystem services.

Available Data and Indicators of Aquatic Biological Condition

To select a complementary metric to the WQI, we limited our search to existing measures of biological condition that are routinely collected at a national scale by the EPA (14, 15). The EPA’s National Aquatic Resource Surveys (NARS) provides a unique set of spatially extensive data, including several ecological indices. As a part of this program, about 2,000 permanent rivers and streams (henceforth streams) and 1,200 lakes (Fig. 1) are sampled every 5 y in collaboration with State and Tribal partners. During both surveys, biological data are collected at each sample location, including counts of fishes (streams), diatoms (streams), benthic macroinvertebrates (streams and lakes), and plankton (lakes). The EPA then derives biological indicators from these data to assess the condition of the Nation’s water resources (https://www.epa.gov/national-aquatic-resource-surveys).

Fig. 1.

The 2013/2014 National Rivers and Streams Assessment (NRSA) and 2007 National Lakes Assessment (NLA) sample sites. The EPA conducts surveys that include new sample locations on a 5-y cycle. NARS ecoregions: WMTNS = Western Mountains, PLNLOW = Plains and Lowlands, EHIGH = Eastern Highlands.

Among candidate metrics developed by NARS, we selected indicators of biological condition that can be quantified on continuous rather than nominal or ordinal scales. We ruled out ordinal descriptions of biological condition (e.g., poor, fair, and good) because estimating benefits of improvements that do not cross a threshold would require assumptions about respondents’ preferences within a category without empirical support. Changes in pollutant levels or other ecosystem stressors were also ruled out because survey respondents would be forced to speculate about the eventual effect of such changes on features of biological condition that they care about. Finally, pollutants may be redundant with indicators included in the WQI, which could lead respondents to conflate the selected metric with direct human use values.

Initial filtering left us with two potential companion metrics to the WQI: 1) multimetric indices (MMI) (16) and 2) observed-to-expected (O/E) taxonomic composition (17, 18). These indicators are both routinely collected and reported by NARS for several thousand streams and lakes across the CONUS. Additionally, both indicators are measured on continuous scales and designed to directly assess the biological condition of aquatic ecosystems. Finally, both have undergone considerable development and testing over the last 40 y (19) and are accepted and widely used in state, regional, and national monitoring programs (20). The detailed mechanics of MMI and O/E approaches are beyond the scope of this paper, although a summary can help clarify the relative advantages of each when assessed with focus groups. Our description of each is tailored to their specific applications within NARS, but numerous variations on these approaches exist.

MMIs are one of the most frequently used tools for assessing the biological health of aquatic ecosystems (21). MMIs are designed to be a comprehensive index of biological condition through the aggregation of several individual metrics that are calculated from taxonomic data (e.g., counts or occurrences of fish, diatoms, benthic macroinvertebrates, or other taxa) (16). Over the decades, hundreds of potential metrics have been proposed for inclusion in MMIs, each designed to capture different aspects of the biological composition or function of sampled aquatic organisms (21). To construct the NARS MMIs, several dozen candidate metrics were assessed following Stoddard et al. (22) to represent six key aspects of biotic condition: 1) habitat preferences, 2) taxonomic richness, 3) pollution tolerance of collected taxa, 4) taxonomic composition, 5) taxonomic evenness/diversity, and 6) feeding groups. Candidate metrics were placed into one of these six categories by experts within the EPA. Once categorized, candidate metrics within each group were subjected to a series of tests. For example, metrics with many zero values or limited ranges were excluded. Further, repeated visits to sample sites were used to identify metrics with stable values to exclude those that are sensitive to sample variation rather than stream conditions. Next, the ability of candidate metrics to distinguish between a set of minimally or least-disturbed sites (also called “reference” sites) (23) and most-disturbed sites was assessed (i.e., index responsiveness). Once a candidate metric was identified to represent one of the six categories, metrics across categories were checked for statistical correlations and uniqueness of biological information to avoid cross-category redundancy. Finally, metric values were scaled to have values of 0 to 10, summed to produce an aggregate index, and rescaled again to 0 to 100 to improve interpretation of scores (22).

In contrast to an MMI, an O/E index quantifies the loss of aquatic taxa due to human-related stressors (24). It does so by comparing the list of taxa observed (O) at an assessed site to the taxonomic composition that would be expected (E) in the absence of stressors (18). To estimate E, a set of minimally or least-disturbed sample sites are identified to serve as regional benchmarks (23). With these reference sites, models are developed to estimate the biological assemblage that would be expected in the highest quality streams or lakes currently extant within a region based on their watershed settings (e.g., watershed area, soils, topography, and climate). Traditionally, O/E assessments used the river invertebrate prediction and classification system (RIVPACS) (17, 18) to estimate the composition of taxa at sites. However, any modeling technique that can accurately predict the probability of taxa occurrences across sites may be used (25). For the EPA’s NARS assessments, E for an assessed site is the sum of individual taxa capture probabilities (P_c), where P_c > 0.5 (17). Likewise, O is the sum of occurrences of taxa at an assessed site, where taxa P_c > 0.5. In this way, O is the count of taxa that in fact occurred from the list of taxa that were expected to occur (i.e., P_c > 0.5) under reference conditions (24). Values of O/E that are <1 indicate the loss of taxa from an ecosystem because E is derived from reference sites (24). However, some variation in O/E values above or below 1 at reference sites is common due to field and laboratory sampling variability, modeling error, and difficulties in identifying high-quality reference sites in some regions to set biological expectations. Typically, O/E values >1 are considered part of this sampling variability and modeling error, although some have attributed O/E values >1 to elevated diversity or shifts in assemblages due to enrichment (26). However, only values of O/E that are <1 are used by the EPA to assess stream and lake conditions as part of NARS.

Selecting an Index of Aquatic Biological Condition

Schultz et al. (12) examined the representation of ecological outcomes in SP studies and identified four criteria that metrics should meet for valid benefit estimation. We used these criteria as a guide to compare MMI and O/E indices. First, ecological indicators used in an SP survey should have measurability. Schultz et al. (12) describe measurability as “a clearly stated relationship to ecological data or model results” to ensure valuation results can be linked to policy outcomes. Subjective descriptions of outcomes, such as “good” or “poor”, often fail to meet this criterion. Second, interpretability ensures that different values of the metric have consistent meanings to survey respondents, subject experts, and resource managers. Third, an indicator with applicability will be relevant to the management scenario that is the subject of the survey. An applicable indicator will aid in scenario acceptance by the SP survey respondent and is required for the SP survey results to be relevant to benefit estimation. Finally, the comprehensiveness of an indicator reflects the degree to which all direct and indirect ecosystem impacts are described by the metric(s).

Our initial selection of the NARS MMI and O/E indicators largely fulfills the criteria of applicability. For a metric to be applicable to the national SP survey and EPA regulatory analysis, it must be available for the entire country, a feature that is unique to the NARS surveys. Further, our selection of metrics on a continuous rather than nominal or ordinal scale is necessary to capture benefits of individual CWA regulations because they tend to have small impacts on water quality that seldom cross the thresholds delineating categorical indicators. These features make both the NARS MMI and O/E indicators salient to benefits analysis of CWA regulations.

The ability of MMI and O/E indicators to assess the biological condition of aquatic ecosystems has been rigorously compared (27) and refined over the last several decades (19). From this work, all aspects of their measurability (e.g., accuracy, precision, sensitivity, and specificity to stressors) improved substantially through refinements in field and laboratory protocols (28, 29) and statistical and other analytical approaches (30). Due to these improvements, MMI and O/E indicators can produce comparable regional assessments of aquatic condition (24). Therefore, we found little difference in measurability between the candidate indicators for use in SP surveys.

MMIs, by definition, are comprehensive measures of biological condition through the inclusion of several life history and behavioral traits (22). The inclusion of the metrics makes them more comprehensive than O/E indices. However, it was unclear how the comprehensiveness of MMIs might affect survey respondents’ ability to separately consider use and existence values when answering the questions. We could not assess the interpretability of MMI or O/E indices through literature alone. To our knowledge, no study has compared the ability of the public to understand or form preferences of biological outcomes based on these indicators. Therefore, the EPA conducted a series of focus groups to evaluate MMI and O/E indices in an SP setting. A total of ten focus groups with eight to ten participants each were conducted in Arlington, VA, Chicago, IL, and Phoenix, AZ. Locations were chosen to work with participants that have a variety of experiences with, and interpretations of, water resource issues. Participants were selected to include equal numbers of men and women with a minimum of a high school diploma and to roughly match the general population regarding race and income. We adopted an emergent design structure (31) for the study, with early focus groups following a conversational format to identify dominant themes when considering environmental quality in aquatic environments. Then, as the study progressed, the discussions became more structured and considered topics such as how to convey scientific information to the public and how to describe changes in our candidate metrics of biological condition.

Each of the focus groups was video recorded, and the recordings were used to generate text transcripts of the conversations. We analyzed the data from each focus group in three sequential steps. First, the observers and facilitator reviewed the video recordings to identify persistent themes regarding the interpretation of each metric. Second, the text transcripts were coded using computer-assisted qualitative analysis software to identify passages referring to each theme. Third, passages were organized by theme and analyzed to draw conclusions about MMI and O/E regarding each theme. For the purposes of choosing between MMI and O/E as our indicator of biological condition, we focused on the anticipated challenges in their application on an SP survey. The challenges that emerged were numerous, and most were shared by MMI and O/E. Two, however, differed in the interpretations of the indicators and influenced our choice.

The first is a conflation of use value and existence value. The ability of respondents to consider changes in the WQI and biological condition separately is critical to our research objectives. It is therefore important for focus group participants to separate the two indicators cognitively when considering changes in water quality. For example, the following passage is taken from a conversation about using MMI to convey biological condition while using WQI to capture suitability for recreational uses (SI Appendix, Fig. S1).

“I assume that any river or lake that’s good enough for me to fish in or swim in is definitely going to support whatever wildlife is in the area. Beyond that, I don’t know. I just assume that if it’s good enough for me, I think that that’s all that matters.”^†

Narrowing the focus of the biological condition indicator to biodiversity of macroinvertebrates or plankton—taxa that are not directly valued for recreation—mitigated the conflation challenge and helped participants to consider the two measures separately when considering willingness to pay for water quality improvements. The following passage is taken from a discussion about using O/E as the indicator to complement WQI as water quality attributes (SI Appendix, Fig. S2):

“I think having two different pieces of information is a good thing. The health of the river is determined by how many species are in there and that kind of thing. Then the other part is how people are using them, so I would say both of them are important.”^‡

The second pivotal challenge is producing a consistent interpretation of environmental changes given an improvement in the numerical value of the indicator. We probed this issue by presenting a ten-point improvement in each of the candidate indicators and asking participants to describe what physical changes would help them to decide whether to vote for a policy that would also increase costs to their household. Participants struggled to describe specific changes when using MMI. Some responded in vague and equivocating ways such as, “Ecological health, that’s a term that might mean different things to different people.”^§ Others responded with questions about what caused the change in the MMI. Upon probing the issue further, we discovered that the construction of the MMI from multiple underlying biotic measures prevented participants from identifying specific and consistent environmental changes that would influence their decision. When asked the same question regarding a ten-point change in O/E, the response from participants was much more direct, “It’s really clear. It’s exactly what it means.”‡ This may be because taxa are a salient “currency” of ecosystems (32) that are conceptually familiar to the public. So, while O/E may capture fewer dimensions of biotic condition, it is one that focus group participants were already familiar with and easily understood.

Our analysis of focus group data regarding the interpretability of indicators of biological integrity favors the choice of O/E to complement WQI on our SP survey. While MMI is more comprehensive than O/E, that feature leads to challenges in interpretation for survey respondents. Given our intention to use the indicator as an attribute on an SP survey, we placed a premium on interpretability. Further, because O/E will be presented alongside the WQI, which captures other elements of water quality, the comprehensiveness of a single measure is less important than if it was the only indication of environmental quality being used.

Focus group discussions also revealed important details on presenting the quantity of resources referenced in the discrete choice questions and the types of taxa represented by O/E. For example, presenting streams in linear units and lakes in areal units was too cognitively burdensome because it divided the resource being valued and doubled the number of attributes respondents must consider when evaluating scenarios. Instead, focus group participants agreed that combined surface area of streams and lakes is an accurate reflection of quantity when assessing water quality improvements and preferred it over other measures. Thus, for the survey, the WQI and O/E values presented in the discrete choice questions will be surface area-weighted averages across all lakes and streams within a region. Another important detail of survey design revealed by focus group research is the importance of emphasizing the use of macroinvertebrates or plankton in biological indices to help respondents focus on motivations unrelated to human uses. Without that emphasis, participants tended to consider taxa with direct use values, such as fish and birds.

Model-Based Interpolation of Resource O/E

Accurate assessment of the public’s willingness to pay for improvements in biotic condition will require accounting for the dependence of preferences on baseline conditions in the watersheds referenced in the SP scenarios and elsewhere (i.e., substitute sites). Currently, available observational datasets lack the spatial density or extent to provide baseline conditions throughout the CONUS. Further, O/E is an internal feature of streams and lakes that cannot be detected through remote sensing. Therefore, we are pursuing modeling to spatially infill between observations and provide the extent and density required by the forthcoming SP survey. For this modeling, we are using an iterative approach that begins with simple regression models of existing NARS O/E scores. This initial modeling will be followed with more complex approaches if needed to achieve the desired accuracy and precision. Here, we present results of these first models that will form the baseline for comparing subsequent efforts.

As response variables, we used available O/E scores of stream benthic macroinvertebrates from the 2013/2014 National Rivers and Streams Assessment (NRSA) (14) and lake plankton from the 2007 National Lakes Assessment (NLA) (15) (SI Appendix, Table S1). These years were used to develop the methodology and establish baseline model performances, but later iterations will include the most current sample years available to provide contemporary estimates of condition for the SP survey. In addition, adjacent years of stream and lake assessments will be used to minimize temporal separation between surveys (e.g., 2017 NLA and 2018/2019 NRSA). A model was created for each waterbody type (i.e., streams or lakes) within each of the three NARS ecoregions for a total of six models (Fig. 1). We created a model for each region because these are the original regions used to model and estimate E for the NARS assessments (33).

To estimate O/E scores at unsampled streams and lakes, the same independent variables must be used for both model calibration and application (34, 35). Therefore, we used watershed metrics from the EPA’s StreamCat (36) and LakeCat (37) datasets (SI Appendix, Table S2). StreamCat and LakeCat contain landscape data for all streams and lakes within the medium resolution (1:100,000) National Hydrography Dataset (NHD) (38). These data characterize both natural (e.g., soils, geology, and climate) and anthropogenic (e.g., urbanization and agriculture) features for the watersheds of 2.65 million stream segments and 378,088 lakes across the CONUS (36). Sample site coordinates were spatially linked with their corresponding streams or lakes to retrieve StreamCat or LakeCat predictor variables for modeling. Because StreamCat and LakeCat contain data for all streams and lakes within the NHD, models can then be applied to unsampled locations to infill between sample sites and produce a continuous map of stream and lake conditions. However, modeled interpolations were only made for 1.1 million stream segments and 297,071 lakes to match the NARS sampling frames (e.g., NARS does not sample ephemeral streams or lakes). Some differences between StreamCat and LakeCat metrics precluded the development of a single model of streams and lakes. In addition, stream and lake biotic assemblages likely respond to different stressors or to the same stressors but in different ways.

As a modeling engine, we used random forest regression. Random forest is a nonparametric machine learning technique that builds many individual decision trees (39) from randomized subsets of the original observations (40). Random forest is advantageous because it makes no assumptions about the normality or independence of input variables. In addition, random forest captures nonlinear relationships and interactions, which other modeling techniques may not. Random forest requires little, if any, tuning, and we used default parameter settings in all models (41). Random forest models have been compared extensively with other machine learning techniques, including such sophisticated techniques as artificial neural nets and boosted regression trees (42–45). Despite their simplicity to implement, random forest often performs as well or better than techniques that require substantial parameterization (42, 44–46), or when compared against linear models (47). Due to these advantages, random forest has become a popular modeling tool across numerous fields, including ecology. During model development, we did not conduct variable (feature) selection because it is unnecessary for random forests when the principal purpose of the model is spatial prediction (48).

To make predictions to unsampled streams, we applied the random forest model to the StreamCat and LakeCat datasets. For continuous values (here, O/E scores), predictor variables are provided to the individual trees of the random forest to produce a prediction for each tree. These values are then averaged across trees to create final predictions for individual streams or lakes across the CONUS (Fig. 2).

Fig. 2.

Distribution of sample sites with O/E scores from the (A) 2013/2014 NRSA and (B) 2007 NLA. Model interpolated O/E scores for NHD (C) streams and (D) lakes. Separate O/E assessments for streams and lakes were developed within three ecoregions: WMTNS = Western Mountains, EHIGH = Eastern Highlands, and PLNLOW = Plains and Lowlands. Dark gray areas in C and D represent streams or lakes that are outside of the sampling frame of the EPA NRSA or NLA, respectively, and were excluded from model interpolation. Despite using the full range of values for modeling, O/E scores >1 are ignored by NRSA and NLA and are truncated to 1 in A–D.

The models explained 25 to 30% of the variation in stream O/E scores and 13 to 36% of the variation in lake O/E scores (pseudo-R²; Table 1). Model root mean squared errors were 0.25 to 0.27 (streams) and 0.22 to 0.25 (lakes) out of observed O/E scores of 0 to 1.6 (Table 1 and SI Appendix, Table S1). Despite explaining a low percentage of the variation in O/E scores, model residual errors showed no spatial biases (i.e., clusters of over or under predictions). To improve the interpretability of mapped O/E scores, we truncated observations >1 because these values are not interpreted by NARS, and doing so is in line with how they will be used for benefits analysis. The models failed to capture the extremes in O/E scores, with both streams and lakes having minimum predicted values of 0.26 and most predicted values occurring above 0.5. However, patterns of higher and lower observed O/E scores in streams (Fig. 2A) and lakes (Fig. 2B) generally corresponded spatially with patterns of low and high predicted values despite not capturing these extremes (Fig. 2 C and D, respectively). Maps of interpolated O/E scores showed distinct shifts in values at ecoregion boundaries (Fig. 2 C and D). For example, marked transitions are visible in interpolated lake O/E scores between the western mountains (WMTNS) and plains and lowlands (PLNLOW) regions in the states of Montana and Wyoming and in streams between eastern highlands (EHIGH) and PLNLOW. These shifts are also present in the original O/E scores yet less apparent due to the lower density of points relative to the interpolated maps (compare Montana lakes at ecoregion boundary in Fig. 2B).

Table 1.

Performances of random forest models of 2013/2014 NRSA and 2007 NLA O/E scores (see Fig. 1 for regional boundaries)

Model	Regions	Psuedo-R2	rmse
NRSA 2013/2014	CONUS	0.28	0.26
	EHIGH	0.25	0.26
	PLNLOW	0.25	0.27
	WMTNS	0.30	0.25
NLA 2007	CONUS	0.30	0.23
	EHIGH	0.13	0.23
	PLNLOW	0.36	0.22
	WMTNS	0.18	0.25

CONUS, Conterminous U.S.; EHIGH, Eastern Highlands; PLNLOW, Plains and Lowlands; and WMTNS, Western Mountains.

Pseudo-R² as defined in Liaw and Wiener (41).

Implications and Future Work

Existence values are an important and large (in aggregate) source of benefits (8) that should be accounted for when considering CWA policies. However, estimating the existence value of aquatic resources presents a major challenge for research economists because of potential overlap and conflation with the use-based values that are the primary focus of the WQI. A further challenge is estimating this complementary metric at a geographic scope that is relevant to resource valuation for national policy. This study advances the state of existence valuation in at least two important ways. First, by comparing candidate metrics with the criteria of Schultz et al. (12) as our guide and with feedback from focus groups, we identified a metric of biotic condition that is interpretable by the public and reduces conflation with the use-based WQI. Second, we described an ongoing effort to model and interpolate biological conditions to unsampled locations to help account for the dependency of survey responses on baseline water quality conditions. Here, we consider these advancements as well as further work that will improve our models of biological condition and link them to other EPA models of water quality.

Selecting a Complementary Metric.

A critical step toward valuation of aquatic resources was identifying a complementary metric to the WQI. Focus groups were invaluable for comparing MMI and O/E indices and refining how we plan to present biological condition to SP survey respondents. For SP study results to be valid, it is critical that respondents consistently interpret and understand the underlying environmental changes that an index is intended to convey. Focus groups also refined our understanding of how different taxonomic groups may influence the interpretation of indices for existence valuation in SP surveys. The use of nonfish taxa, such as macroinvertebrates and plankton, helped respondents consider ecosystems and management scenarios independently of their impacts on used-based values, such as recreation and natural food production. This finding is important as the analysis of surface water regulations expands beyond human health and direct use benefits. Additionally, due to their diversity and responsiveness to stressors, benthic macroinvertebrates are one of the most used taxonomic groups in freshwater bioassessments worldwide (19, 49), making them good candidates for nonuse benefits studies globally.

Modeling Resource Condition.

Currently, the models explain a low proportion of the variation in O/E and exhibit marked transitions in O/E scores along ecoregion borders. The poor model performance is likely due, in part, to the fact that values of E in the O/E ratio are estimated from previous models that include model error. Thus, there is likely an upper limit to how well models based on the existing NARS O/E scores can perform. The hard transitions in O/E scores at ecoregion borders occur because reference sites in the PLNLOW ecoregion are rarer and are often of lower quality than the WMTNS and EHIGH ecoregions (35). Thus, when the O/E assessments were constructed, the regional models that predict E were derived from, and predict, different baseline conditions. Therefore, it should be emphasized that the ecoregional shifts observed in our interpolated values are not an artifact of the models constructed here. Instead, the spatial distribution of modeled values appears to accurately reflect the calibration data (Fig. 2), and mapping these shifts at ecoregion boundaries simply highlights the extent of this issue. Further, it emphasizes the challenge faced by national assessments to identify high-quality reference sites in regions with extensive land use (50, 51) and the need to develop bioassessment methods that do not rely on regional reference sites (52). Because these shifts are inherent in the calibration data, adjustments to the current maps, such as smoothing at regional borders, may obscure these transitions but will not solve the underlying problem, especially toward region centers. Instead, as of this writing, we are pursuing modeling that does not use the O/E scores provided by NARS. Rather, we returned to the original taxonomic data of the NARS samples to construct CONUS-scale models (i.e., inclusive of all sites across ecoregions) that predict the distribution of individual taxa nationally and, hence, the biotic assemblage (25). Although preliminary, models of individual stream benthic macroinvertebrates constructed at the national scale do not suffer from harsh boundary transitions and outperform regional-scale models of the same taxa. In addition, these initial models also appear to outperform the RIVPACS models used to produce the original NARS assessment. These preliminary models show promise for providing the coverage, resolution, and accuracy required to support the EPA’s benefits analysis of CWA regulation, but additional testing and validation will be required.

Despite limitations of the current models, this study shows promise for providing interpolated estimates of ecological condition to complement the WQI. The models of O/E presented here, and under development, illustrate how a continuous measure of biological condition can be spatially interpolated for use in resource valuation, such as our forthcoming SP survey. The fine spatial resolution and near-continental extent of the interpolations will have several advantages for SP survey analysis. First, this framework is flexible. Analysis of streams and lakes in tandem or in isolation is possible because each resource type was modeled separately. Second, interpolations can be spatially partitioned or aggregated to fit a variety of needs, such as valuation of resources within a political or management boundary or to match other EPA models and tools used for aquatic resource assessment and valuation, such as the EPA’s Hydrologic and Water Quality System (HAWQS) (53) and Benefits Spatial Platform for Aggregating Socioeconomics and H₂O Quality (54). However, care must be taken to understand how the dominance of a resource type might vary as interpolated values are aggregated. Aggregation to coarser resolutions may weight lakes over streams due to their areal dominance at larger scales when equal or alternative weighting may be desired. Yet the flexibility of the framework will allow examination of interpolated O/E scores at their original resolution to understand how aggregation will affect analytical results.

Linking Policy Scenarios with Biological Condition.

The purpose of this research is to provide resource economists with the data and tools they need to estimate changes in value in response to proposed management. It is unlikely that the models of O/E described here can provide this capability. Rather, their purpose is to interpolate current conditions to unsampled lakes and streams to account for geographic variation in benefits valuation. The large number of watershed variables we used in the modeling (SI Appendix, Table S2) make it difficult to infer the mechanistic pathways by which proposed policies would affect biological condition through improvements in water quality or habitat condition. Instead, resource economists will need the means to forecast biological condition under baseline and regulatory scenarios. However, substantial work remains to build such forecasting capability. First, explicit linkages must be made between candidate policy options and human-related watershed activity, instream water or habitat quality, and O/E scores. For example, to infer potential sources of impairment in Nevada streams, Vander Laan et al. (55) related macroinvertebrate O/E scores to measurements of heavy metals and observed total dissolved solids (TDS) relative to TDS values expected under reference conditions. Values of this TDS O/E were largely driven by agriculture, mining, and urbanization, thereby providing an indirect assessment of how major land uses influence biological condition through water quality. Second, when possible, it will be important to use modeling approaches that can describe causal linkages between management and biological condition. Schmidt et al. (56) used structural equation modeling to link land use in Midwestern watersheds to instream habitat structure, temperature, pesticide and nutrient concentrations, and, in turn, algal and benthic macroinvertebrate MMIs. Such models could disentangle mechanistic pathways between human-related landscape stressors, in situ water and habitat quality, and O/E scores. Defining these pathways will be challenging given the many direct and indirect ways human activity can affect aquatic biota. Additionally, as Vander Laan et al. (55) showed, important linkages can be identified using less formalized techniques than structural equation models. Finally, the models described here are spatial models; that is, the relationships explain differences among sample locations rather than changes due to management over time. Although there is some evidence to support space-for-time substitution in ecological models (57), such assumptions should be explicitly tested.

Process-based models could be used to relate management scenarios to water quality parameters. This capability exists through water quality assessment tools such as the EPA’s HAWQS model (53). Such models allow a researcher to enter location-specific changes in pollutant loadings and/or management practices and recover the downstream water quality values needed to calculate the WQI. However, further work is needed to link these water quality outputs with outcomes of O/E. At present, process-based models are difficult to implement at a national scale, but recent advances in large-scale modeling (e.g., National Water Model; https://water.noaa.gov/) suggest that such capabilities are forthcoming. Finally, despite the process-based nature of these water quality models, the linkages to O/E values would still need to be made through spatial models due to the relative rarity of long-term biological datasets.

Despite these challenges in modeling O/E in policy scenarios, the incremental advancements described here can improve the way EPA values water quality changes. Analysis of focus group data revealed that indicators that are aggregations of multiple underlying metrics pose challenges to interpretability for SP survey respondents. Further we found that presenting O/E as a measure of biodiversity for taxonomic groups with no direct use value is helpful for respondents when cognitive separability between use values and existence values is necessary for research objectives. Estimating use values and existence values in a separable way will allow more accurate benefits estimation for scenarios in which the quality of recreational experiences and biological condition are not perfectly correlated. Further, given that there is no economic justification for existence value to decline as distance from the resource increases, the forthcoming SP survey will provide empirical support for decisions regarding distance decay and the extent of market for nonuse values. Finally, modeling the O/E scores of lakes and streams that have not been sampled will allow the EPA to estimate willingness-to-pay functions that depend on baseline conditions anywhere in the CONUS. If subsequent models are successful, a national dataset of stream and lake O/E scores could provide resource economists with a consistent index of biological condition for benefits valuation research and improve comparability among studies.

A final consideration for future work will be addressing how climate change is shifting water quality baselines and the consequences for benefits estimation. Climate change is warming watersheds and altering precipitation patterns across the United States (58). These shifts will trigger wide-ranging changes in aquatic ecosystems, including severe consequences for aquatic biota (59). These changes will also impact the tools we use to assess biotic condition and conduct benefits estimation. Thus, future work must address these impacts on biotic integrity and how we value this integrity. Much like the work to link policy scenarios to biological condition, this work will need to model how scenarios of climate change will translate to the hydrologic, chemical, and thermal regimes of aquatic ecosystems and the consequences for biota. Further, to our knowledge, little work has been done to assess the public’s willingness to pay to avoid climate change–related loss in biological integrity, especially in a nonuse context. Such work will be important to avoid underestimation of these benefits and provide critical insight for decision makers when considering CWA policies under a changing climate.

Concluding Remarks

We reported on the results of the first two phases of a multiphase project to improve the way the EPA values aquatic resources: 1) identify a metric that satisfies important criteria for use in SP valuation and 2) interpolate observational data to achieve the spatial resolution required to estimate the valuation equation. Focus groups were critical for evaluating indices of aquatic biological condition and exposing a tension between index comprehensiveness and interpretability. In the end, concerns of interpretability outweighed those of comprehensiveness, and we found that O/E, while not perfect, satisfies the criteria best and suits our specific purpose of valuing water quality changes throughout CONUS. We also reported on modeling to interpolate O/E values to streams and lakes across CONUS. Our hope is that spatially explicit maps of ecological condition can improve nonuse benefits estimation based on SP surveys. Although additional refinements are needed to improve model performance, a nationally consistent dataset of biological condition could facilitate the acquisition of such data by research economists and help to improve comparability among studies of nonuse benefits, especially benefits associated with existence value.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Previously published data were used for this work: US EPA Office of Water and Office of Research and Development (2020) NRSA 2013/2014: A collaborative survey (EPA 841-R-19-001), Washington, DC (14); US EPA Office of Water and Office of Research and Development (2009) National Lakes Assessment 2007: A collaborative survey (EPA/841/R-09/001), Washington, DC (15) (https://www.epa.gov/national-aquatic-resource-surveys); R. A. Hill, M. H. Weber, S. G. Leibowitz, A. R. Olsen, D. J. Thornbrugh, The Stream-Catchment (StreamCat) dataset: A database of watershed metrics for the conterminous United States. J. Am. Water Resour. Assoc. 52, 120–128 (2016) (36) (https://www.epa.gov/national-aquatic-resource-surveys/streamcat-dataset); R. A. Hill, M. H. Weber, R. M. Debbout, S. G. Leibowitz, A. R. Olsen, The Lake-Catchment (LakeCat) Dataset: Characterizing landscape features for lake basins within the conterminous USA. Freshw. Sci. 37, 208–221 (2018) (37) (https://www.epa.gov/national-aquatic-resource-surveys/lakecat-dataset). All other data are included in the manuscript and/or SI Appendix.