Note: This page provides links to non-EPA websites that provide additional information about this topic. You will leave the EPA.gov domain, and EPA cannot attest to the accuracy of information on that non-EPA page. Providing links to a non-EPA Web site is not an endorsement of the other site or the information it contains by EPA or any of its employees. Also, be aware that the privacy protection provided on the EPA.gov domain (Refer to Privacy and Security Notice) may not be available at the external link. The following links exit the site Exit
A range of factors—including regulatory priorities, spatial scales, conceptual models, data quantity and quality, and resource constraints—should be considered in selecting the analytical approach, or combination of approaches, to use for developing numeric nutrient criteria (NNC). These factors are often structured as a sequence of decisions that join management and policy goals, technical capacity, and resource constraints. This section explores these factors in detail and offers specific examples of how they can affect the decision to apply the reference condition, stressor-response, or mechanistic modeling approach to criteria development. Analysts also can consider applying more than one approach as an alternate way of independently analyzing the same data. Applying a second approach can often corroborate analytical results from the first approach and strengthen decisions on estimating criteria.
The Clean Water Act directs state and tribal regulatory agencies to develop new or revised surface water quality standards regulations (Title 33 United States Code [U.S.C.] § 1313(c)). In carrying out these responsibilities, the agencies sometimes develop priorities and goals focused on developing NNC (e.g., for lakes, streams, and cold-water fisheries). They can be expressed in a variety of ways, including:
- in state nutrient criteria plans (e.g., the Mississippi Department of Environmental Quality nutrient criteria plan Exit),
- in state 305(b) reports pursuant to CWA section 106 (e.g., the Alabama Department of Environmental Management 2014 Integrated Report), and
- through collaborative coordination with the U.S. Environmental Protection Agency (e.g., the Mississippi River Gulf of Mexico Hypoxia Task Force).
States and tribal regulatory agencies might express their priorities spatially, intending to develop NNC for different groupings of water bodies ranging from site-specific waters to entire classes of water body types (e.g., headwaters and cold-water streams). Timeframes for developing NNC might coincide with regular triennial reviews or be motivated by other regulatory interests such as a National Pollutant Discharge Elimination System (NPDES) permit being issued or total maximum daily load (TMDL) development. Constrained regulatory timeframes might favor the reference condition or stressor-response approach over mechanistic modeling, which can have greater data and model development demands. Finally, state narrative nutrient criteria often contain language that seeks to limit unwanted or adverse environmental impacts by minimizing nutrient pollution.
Examples of state narrative nutrient criteria for Idaho and New Mexico:
Surface waters of the state shall be free from excess nutrients that can cause visible slime growths or other nuisance aquatic growths that impair designated beneficial uses. (IDAPA 58, Title 01, Chapter 02, Section 200.06, “Excess Nutrients” Exit )
Plant nutrients from other than natural causes shall not be present in concentrations which will produce undesirable aquatic life or result in a dominance of nuisance species in surface waters of the state. (NMAC, Title 20, Chapter 6, Part 4, 22.214.171.124.E, “Plant Nutrients”)
This language can motivate the regulatory agency to pursue a stressor-response approach, using empiricism to illustrate connections between nutrient pollution and adverse environmental effects implied in state’s narrative nutrient criteria. Reference approaches and mechanistic modeling also can be applied to estimating nutrient pollution levels to prevent adverse environmental effects specifically mentioned in the state’s criteria. Surface waters free from or having minimal human activity (i.e., reference quality) and having no adverse environmental effects can be modeled to establish nutrient pollution expectations for similar water bodies in the state.
If time and resources are not limited, site-specific regulatory action on a well-studied water body might favor a mechanistic modeling approach or a combination of approaches. The Montana Department of Environmental Quality developed a numerical simulation model Exit for the Yellowstone River as a way to simulate nutrient pollution impacts. Broad regulatory action on an entire class of state waters consisting of many distinct water bodies might necessitate a reference and/or stressor-response approach from which broad water quality inferences can be made. The Florida Department of Environmental Protection developed nutrient criteria for all the state’s class III lakes using a classification system based on lake color and alkalinity (FDEP 2012 Exit, Chapter 10.3).
A conceptual model can predict the effects of nutrient pollution on a water body’s designated uses. It can explicitly express the many factors that affect or modulate nutrient pollution in the aquatic environment. Knowledge of these factors informs both the data collected and the data used to construct reference distributions, stressor-response models, and mechanistic models. Conceptual models also offer a way for state and tribal agencies to communicate clearly to stakeholders the designated uses that will be protected and through which ecological or biological endpoints protection will be conferred. Conceptual models may help inform which approach is best linked to and protective of management goals.
The amount of data available is a critical factor guiding state and tribal decision-making on the analytical approach to use. NNC development at its core is an analytical endeavor and a sufficient amount of data is needed to construct rigorous models and accurately estimate the protective water quality conditions. Data sets that contain a limited amount of spatial and/or temporal data might be able to produce reference, stressor-response, or mechanistic models, but the results will have greater explanatory variability and prediction uncertainty than data sets with more spatial and/or temporal data. Data sets that are highly resolved in time and space can benefit all approaches and often yield more accurate model estimations and predictions of water quality. In cases in which data quantity is sufficient, but the range of water quality parameters is constrained, model development can be limited using any of the three approaches because the appropriate parameters are not available to reduce variability in the models. For example, a 40-lake data set could have 12 lake TN and TP measurements per year for 10 years, but no additional data that describe important lake properties such as depth, area, and elevation. The result would be reduced variability in a reference distribution model from which potential TN and TP criteria could be estimated. Going a step further, lack of a biological response variable indicative of designated use protection (e.g., algal biomass as chlorophyll a) would preclude employing a stressor-response approach despite having ample data on the stressors, TN and TP.
Data Quantity—Screening Effects
Data screening is a common practice associated with the early stages of all three analytical approaches and is used to create customized reference data sets and models that exhibit specific characteristics of interest (e.g., streams coincident with more than 95 percent forested watersheds, shallow lakes less than 10 meters deep, or summer epilimnetic chlorophyll a). Discovering data that fit certain characteristics can assist in identifying a reference population or yield unique stressor-response models. The trade-off is that the original data set can be reduced to the point at which statistical estimates (e.g., 90th percentile, confidence limits around the mean linear regression line) are more uncertain, undermining confidence in making inferences about the water quality expectations of a broader population of waters.
Data quality and integrity also can affect data quantity. Data collected under strict data quality objectives can provide confidence in subsequent model development. Conversely, poor or imprecise sample collection methods and analytical processes, or those of unknown quality, can limit a data set and the subsequent analysis. For example, samples that have been improperly handled (e.g., lack of refrigeration or sample preservation) might yield analytical results that either overestimate or underestimate the true analyte concentration. In addition, less sensitive instruments might yield measurements that, while precise, are not accurate, overestimate true analyte concentration, and may be unable to resolve concentrations at which adverse responses occur. This, in turn, can yield heavily censored data sets resulting in biased descriptive statistics and models. Close attention to the underlying data quality objectives (e.g., quality assurance project plans) and inspection of the data sets (e.g., outliers, detection limits) can strengthen both the choice of approaches and the interpretations of the results derived from those approaches.
Developing NNC can consume significant state or tribal staff time, involve specialized technical skills, and incur monetary costs associated with data management and analysis. Staff availability and technical expertise, in combination with factors described above, may weigh heavily in the decision to pursue a particular analytical approach:
- Mechanistic models often require specialized expertise in working with water quality simulation models and their components, including model domain, calibration, and model outputs.
- The stressor-response approach might require expertise in using statistical-specific analytical software (e.g., R or SAS).
- The reference condition approach often requires knowledge of the scientific literature published on analytically defining conditions associated with a reference condition (e.g., Hughes et al. 1986).
States and tribes must consider their existing technical capabilities, the need for additional technical capacity, and the cost of acquiring that capacity in determining which approach(es) to pursue.
Approach-Specific Selection Factors
Each of the factors discussed should be considered in determining whether any of the approaches is appropriate for developing NNC. Data quantity and quality, for example, are paramount to constructing representative reference distributions, accurate empirical models, and deterministic models that reflect the observed environmental conditions. In addition, factors unique to each approach also should be considered when weighing the various approaches, including the differences in the spatial and temporal scales of the outputs that each approach yields.
Reference Condition-Specific Factors
A key underlying assumption with the reference condition approach is the explicit choice made regarding what the reference water quality condition represents—natural or pristine water quality conditions, least-disturbed (“the best of what remains”), or a more impacted water quality state. Initially, a natural water quality condition may be desired, yet data availability (e.g., space, time, quality, quantity) may constrain the analyst’s ability to estimate that desired condition. A stressor-response approach might be the better choice for inferring where natural water quality conditions lie along the gradient of the available water quality data. If data constraints are tied to a lack of broad spatial coverage in the data set, estimating natural water quality conditions might be overcome if data are available that make a temporally based estimate (or historical estimate) of the natural water quality conditions possible (refer to table 1 in the Reference Condition section for a summary of the temporal and spatial reference condition approaches).
Another key assumption is the strength of the companion data that corroborates the desired water quality state of the reference population. Data that indicate minimal physical anthropogenic impacts (e.g., high habitat quality scores, forested land use) are strong indicators that reference conditions might be present. Biological data associated with natural or least-impacted water quality conditions also are often leveraged to substantiate those assumptions. Like the reference condition approach itself, the data quantity and quality associated with the corroborating biological data are equally important. These data often relate directly to designated use attainment, and biological data in particular can be used to infer attainment of aquatic life uses.
The strengths of the stressor-response approach are its ability to reflect known pathways and associations between nutrient pollution and environmental effects, including effects on assessment endpoints linked to management goals. The strength of the inferences made using this approach, therefore, hinge on the quality of the conceptual models upon which the empirical models are based. The conceptual models, in turn, can promote detailed expressions of the measures of effect and exposure that link directly to the designated uses the NNC will protect. Once again, quantity and quality of the accompanying data for which measures of effect and exposure can be estimated are fundamental to this approach.
The ecological context in which stressor-response models exist is an important selection factor, too. Nutrient pollution models of oligotrophic waters may not yield accurate estimates of the effects because the measure of effect (e.g., algal biomass) might be weak in this ecological context or the nutrient stressor data might be constrained to a small gradient, resulting in an abbreviated model. Conversely, models that reflect eutrophic conditions might be constrained in estimating less productive conditions that are protective of designated uses. In such cases, protective conditions might have to be estimated using extrapolation techniques (USEPA 2010c). Classification variables and their associated data also are important in accounting for the variation in stressor-response models. When present and integrated into stressor-response models, estimates on the measures of the effect and exposure become more accurate and credible for regulation.
Mechanistic Model-Specific Factors
The selection factors that apply specifically to the stressor-response model approach also can apply to and have cascading impacts on the mechanistic modeling approach. Water quality simulation models use mathematical equations describing system behavior that can predict outcomes based on component variables. These equations, particularly for biological processes in water quality simulation models (e.g., algal nutrient uptake, growth rate, biomass accumulation rate, and oxygen production and consumption rates), are based on empirical data, including the results of stressor-response models. Thus, the strength of a mechanistic model is in large part dependent on the strength and accuracy of the underlying stressor-response models that drive the mechanistic model. The measures of effect and exposure, as in stressor-response models, should have firm foundations in known conceptual models and be made explicit when evaluating potential protective water quality conditions.
The outputs of mechanistic models are often time-variable, which makes interpreting output from these models uniquely different from interpreting reference condition distribution statistics or statistics associated with stressor-response models. For example, a three-dimensional water quality simulation model such as WASP can yield nutrient concentration predictions every 5 minutes for hundreds of model cells covering area and depth. These model cell predictions occur at timescales many orders of magnitude shorter than predictions made using annualized data in a reference condition distribution or even a seasonal (e.g., growing season) stressor-response model using nutrients and algal biomass as the response variable. Also, a transition will be necessary from the short timescales of mechanistic model output to the longer timescales embodied in a nutrient criterion’s magnitude, duration, and frequency.
Selecting the Analytical Approach
After weighing each of the factors discussed, the state or tribal agency must select the approach that:
- Accommodates the state regulatory constraints as well as the timeframe, resources, and technical capabilities;
- Maximizes participation of the available data in the analysis; and
- Provides rigor to and confidence in the analytical results.
Deciding to use more than one approach has the potential to corroborate analytical results and strengthen the choices made in estimating criteria. It is common for states to select reference condition and stressor-response models as approaches for deriving and estimating criteria, especially for streams and lakes across broad regions. The two approaches often rely on the same data, the costs and technical expertise associated with developing either models are typically low to moderate, and the approaches complement each other in terms of the inferences made about designated use protection from the results (i.e., designated use protection is implicitly inferred from reference models, whereas it is explicitly inferred from stressor-response models). Table 1 provides a comparison of selection factors among analysis approaches.
Table 1. Quick tips comparing technical approaches
|Selection Factors||Reference models (spatial)||Reference models (temporal)||Stressor-response models||Mechanistic models|
|Frequency of use||Common||Rare||Common||Rare|
|Link to designated use protection||Implicit||Implicit||Explicit||Explicit|
|Spatial resolution of outputs4||Water body||Water body||Water body||Fine-scale to water body|
|Temporal resolution of outputs5||Seasonal to annual||Seasonal to annual||Seasonal to annual||Minutes to interannual|
|Water quality data inputs||Moderate||Moderate||Moderate||High|
|Auxiliary data inputs6||Moderate||Moderate||Moderate||High|
|Downstream protection estimates||Limited7||Limited7||Limited7||Yes8|
|Technical expertise||Low||Low||Moderate to high||High|
|Expense||Low||Low||Moderate to high||Moderate to high|
|1 Large populations of waters at the continental, regional, state, or watershed scales.
2 Uses the historical record of individual water bodies; thus their application is usually limited to site-specific applications.
3 Site-specific waters or multiple waters that are hydrologically connected.
4 Reference model outputs (statistical estimations) and stressor-response model outputs (predictions) are often used to infer water body-wide conditions. Mechanistic model outputs (predictions) can simulate spatially explicit conditions in a water body (e.g., surface, bottom, nearshore, open water) or be used to infer water body-wide conditions. Advances in water quality sensor technology (e.g., satellite remote sensing) are enabling more highly resolved spatial observations.
5 Reference model outputs (statistical estimations) and stressor-response model outputs (predictions) typically infer aquatic conditions that occur on the seasonal or annual timescale. The scales of the model output reflect the scales of the data, which are often generated from discrete monthly or quarterly ambient environmental samples that are then averaged at seasonal or annual timescales. While mechanistic models often use data obtained at the same timescales as stressor-response models, computer simulations can be conducted to make finer scale predictions in time (e.g., every 5 minutes, hourly, daily, or weekly) dependent on computing power and project time constraints. Model outputs can be aggregated and scaled up to longer timescales (e.g., seasonal, annual, or interannual). Advances in water quality sensor technology (e.g., in situ nutrient sensors) are enabling finer scale temporal observations ranging from minutes to weeks.
6 In addition to water quality data, auxiliary data are often used in model development to quantitatively characterize reference site quality, to classify waters to reduce statistical variation in model distributions or predictions, or to simulate physical dynamics within which mechanistic model simulations of biological processes occur. Examples include land use, lake area, depth, elevation, alkalinity, and water clarity; stream flow and gradient; and estuary salinity and temperature.
7 These approaches can be used to estimate the water quality conditions for upstream waters that protect downstream waters provided that the models are constructed to reflect their hydrologic connectivity.
8 This assumes the downstream water body model domain extends upstream into the tributary system(s) of interest.