The peer-reviewed literature documents clear relationships between human transformation of the natural world and increased risk of disease (Plowright et al. 2008; Myers and Patz 2009), emphasizing the need for more structured ecosystem service assessment approaches (Holzman 2012). The potential impacts of failing to act on sustainability in a significant way are highlighted by Baronosky et al. (2012), who demonstrate that a planetary-scale critical “tipping point” as a result of human influence is fast approaching, the consequences of which are not to be underestimated. If the goal of science and society is to guide the biosphere toward desirable conditions, as opposed to those thrust on us unwittingly, Baronosky et al. (2012) argue unequivocally that the time to act is now.
However, at different scales, there are actually multiple, interactive tipping points and exceeding one is likely to alter the threshold of others (Cairns 2009), leading to a system-wide cascade of potential impacts and complicating efforts to evaluate causal relationships. Because ecosystem services frame environmental benefits in terms of human well-being, the hope, if not expectation, is that incorporating this concept into decision making will lead to sustainable policy development and implementation (e.g., recognizing the importance of natural capital and green infrastructure necessary to generate valued outcomes). However, as sustainability (that is defined and interpreted differently depending on the context) becomes both a goal and a metric against which to evaluate decisions, increasingly complex information must be incorporated into the decision-making process (Levin and Clark 2010). Ecosystem services, sustainability, and tipping points are mediated through complex and large-scale processes that are not amenable to traditional reductionist approaches to causal inference (Plowright et al. 2008) and call for sets of models and statistical tools that transition analysts from experiments that detail localized biological processes to landscape-scale patterns where management and policy take place (Cardinale et al. 2012). By definition, assessments that are based on sustainability as a guiding principle require approaches that go beyond individual biophysical processes.
The context-dependent definition and interpretation of sustainability creates challenges for effective implementation of management approaches defined as “sustainable.” Take agriculture as an example. Proponents of large-scale, mechanized farming with use of biotechnology and agrochemicals will argue the sustainability of that approach, whereas small-scale, labor-intensive, diversified farmers will argue the same; the question becomes, how is sustainability being defined and what are the metrics by which sustainability as a goal will be gauged? These specific definitions emerge from the decision-making context.
Further regulatory and technical constraints exist at more refined scales. Regulatory frameworks provide the basis for most site-specific decisions, particularly in the United States, and constraints follow from guidance that exists for particular processes that make it challenging to explicitly incorporate ecosystem service endpoints, for example, the human health and ecological risk assessment process. The Comprehensive Environmental Response, Compensation, and Liability Act (CERCLA), commonly known as Superfund, was enacted by Congress on December 11, 1980, and uses a risk-based approach to calculating clean-up levels at Superfund sites. Similarly the Resource Conservation and Recovery Act (RCRA 42 U.S.C. §6901 et seq. 1976) also follows Superfund risk assessment guidance. Natural Resource Damage Assessments (NRDA) have their own guidance for injury determination (NOAA 1996), and the overlap between NRDA and risk-based frameworks is a topic of discussion in and of itself. In general, most regulatory programs focus on particular approaches and metrics that are not necessarily based on sustainability principles, or it is difficult to include those explicitly.
The clearest technical constraints lie in developing quantitative relationships between ecosystem services of interest and the ecological functions and processes that give rise to those services, and then the subsequent economic valuation—if not monetization—of those ecosystem services. Although ecological models are the tools most suited to dynamically capturing changes in service provision that can be linked with their economic counterparts for testing alternative scenarios of impacts of human activities (Galic et al. 2012), there are always multiple, potentially conflicting goals for the use of natural resources or for what are the “best” uses for an ecosystem and associated habitats. Resiliency and evolutionary adaptation in nature are particularly difficult to understand and model and represent a fundamental aspect of the observed complexity in attempts to model ecological systems (von Stackelberg et al. 2007).
Decision analytic techniques represent a set of tools that provide a structured framework for explicitly evaluating potential tradeoffs with respect to defined criteria across management alternatives. Sustainability as a goal or management objective can be defined as a set of strategic outcomes within a particular context (e.g., remediation of a contaminated site considering alternatives for future land use and how remediation will be conducted). The evaluation criteria provide the specific metrics against which the sustainability of alternatives will be gauged (e.g., energy use, risk-based metrics for remediation outcomes, climate impact, and so on). The process of identifying the alternatives and developing criteria and criteria weights against which to evaluate each of the alternatives can be as important as the ultimate ranking of alternatives. The interaction, communication, and identification of key issues involved in developing the decision analytic approach is an opportunity to frame the relevant issues across stakeholders, and through developing criteria and their associated weights, identify the issues of importance to those stakeholders. The process can be more important than the outcome.
For many, multicriteria decision analysis (MCDA) constitutes the whole of decision analytic techniques, when in fact there are not only numerous MCDA methods (Kiker et al. 2005), but a range of other approaches that are properly considered as part of decision analysis (Clemen and Reilly 2003). A recent US Environmental Protection Agency (USEPA) Board of Scientific Counselors report highlights several different decision analytic methods and the ways in which they can be applied (USEPA, BOSC, 2010).
Ecosystem services and the risk assessment process
Guidance for ecological risk assessment has been formally in use since the early 1990s (USEPA RAF 1992, 1998; USEPA 1997) and is generally acknowledged to “stand the test of time” with respect to providing a useful framework for decision making (USEPA SAB 2007; Dale et al. 2008). However, in practice, risk-based decisions that rely on strict numerical thresholds for contaminant concentrations typically fail to account for the ecology surrounding those concentrations (Galic et al. 2012).
It is possible to develop numerical thresholds in the context of ecosystem services as proposed by Nienstedt et al. (2012). They define 7 key drivers for ecosystem services (microbes, algae, nontarget plants [aquatic and terrestrial], aquatic invertebrates, terrestrial nontarget arthropods including honeybees, terrestrial nonarthropod invertebrates, and vertebrates) related to the potential impact of pesticide use. Rather than developing simple numerical ecotoxicological values, the authors address 6 dimensions, including: biological entity, attribute, magnitude, temporal and geographical scale of the effect, and the degree of certainty that the specified level of effect will not be exceeded (Nienstedt et al. 2012).
Numerical thresholds may be appropriate in the context of chemical use (e.g., agricultural use of pesticides), but many decisions, particularly where chemicals have already been released (e.g., remediation) are less amenable to a strict reliance on risk-based levels. Chemicals are not the only stressors in the environment, and that fact forms a significant part of the rationale underlying ecosystem services.
In fact, the most significant stressor faced by ecosystems generally and ecological receptors in particular is loss of habitat, given that maintenance of healthy and sufficient habitat is a necessary precondition for the provision of all ecosystem goods and services, directly or indirectly (de Groot et al. 2002). The relationship between habitat and biodiversity is well-established (Brooks et al. 2002; Graham et al. 2006) across both terrestrial and aquatic environments (Gratwicke and Speight 2005). Furthermore, the ecosystem consequences of local species loss are as quantitatively significant as the direct effect of several global change stressors (Hooper et al. 2012). There would seem to be only a minor role for toxicological thresholds in that context.
In and of itself, however, ecological risk assessment as a framework is well-suited to move beyond a single-stressor (e.g., chemical) focus. Indeed, the very first sentence from USEPA guidance states “ecological risk assessment is a process that evaluates the likelihood that adverse ecological effects may occur or are occurring as a result of exposure to one or more stressors,” (USEPA RAF 1992, 1998). The guidance goes on to define assessment endpoints as the explicit expression of the environmental values that are to be protected, operationally defined by ecological entities and associated attributes (USEPA RAF 1998). In the context of the guidelines, an environmental value refers to a component of the environment valued by society: in other words, ecosystem services, even if not formally defined in that way. A subsequent USEPA document on developing generic assessment endpoints does specifically identify the environmental “values” associated with assessment endpoints in ecosystem service terms as shown in Table 1.
Table 1. Environmental values for assessment endpoints
|Consumptive||Direct commodity value: food, energy, timber, fiber, pharmaceutical and industrial products|
|Functional||Ecological functions benefitting public health and welfare: pollen and seed dispersal, moderation of weather extremes, waste detoxification|
|Recreational||Recreational opportunities such as fishing, boating, hiking; can be passive use, or direct economic activity (e.g., tourism)|
|Educational||Natural and scientific study|
|Option||Value to future generations of preserving the option of using the environment at some future time|
|Existence||Value of existence of ecological systems independent of direct service or function: aesthetic, moral, cultural, religious, spiritual|
USEPA does not identify habitat as a generic assessment endpoint for anything other than threatened and endangered species because it is the organisms that are valued directly, whereas by definition habitat is that which supports organisms and thus is valued indirectly (USEPA RAF 2003). However, habitat can be a useful proxy when considering the impact of multiple stressors, particularly those that impact habitat directly (NOAA 2006). If the ecosystem service outcome to be maximized is defined as a sustainable population of a particular species in a prospective assessment, habitat represents the most appropriate assessment endpoint given the potential for human development to be a greater stressor than chemical contamination. In a retrospective assessment, the assessment endpoint of the organisms directly rather than habitat can lead to a situation in which the condition of the species themselves is emphasized (e.g., estimating population size and the relationship to chemical exposures) with potentially little motivation to determine why the species may be lacking.
The policy interpretation of risk assessment rather than the risk assessment process itself appears to emphasize specific numeric thresholds. The USEPA is currently developing generic assessment endpoints that are described directly in ecosystem service terms (see the proposed action plan developed by USEPA's Risk Assessment Forum in response to the USEPA Ecological Processes and Effects Committee [USEPA SAB 2007] report [Suter and Maciorowski 2012; USEPA RAF 2010]). Assuming the guidance is accompanied by direct policy implementation and interpretation, this may lead to greater adoption of ecosystem service assessments in an ecological risk assessment context.
Decision analytic approaches for integrating risk and ecosystem services
A framework for considering both ecosystem services and the results of risk assessments, as well as other modeling results and disparate endpoints, is through the use of decision analytic frameworks (Liu et al. 2010; Turner et al. 2010). The 2 methods of most use with respect to ecosystem services include influence diagrams and MCDA, and in both cases, explicit links to GIS or other spatially explicit visualization techniques enhance the analytical capabilities of the methods.
Influence diagrams, often referred to as Bayesian networks (BNs), or as a subset of Bayesian networks, are models that graphically and probabilistically represent relationships among components of a conceptual model (Barton et al. 2012). A recent issue of Integrated Environmental Assessment and Management (IEAM 8(3), July 2012) provides several examples of the ways in which BNs can better inform environmental management decisions by integrating different kinds of relationships (e.g., socioeconomic, ecological) in one framework. Schultz et al. (2011) also provide a summary of recent BNs from the literature. The graphical construction of influence diagrams allows diverse stakeholders to visualize relationships across model components and to explore management questions. When linked with GIS, this approach becomes particularly useful in the context of ecosystem services, as all services are ultimately linked to land use changes and human activities on a landscape scale (Karvera et al. 2011; Swetnam et al. 2011). Arguably, the most widely used tool for ecosystem service assessment, InVEST, is fundamentally GIS-based (Tallis and Polasky 2011). Understanding the impacts of human activities on the environment from a local to a global scale requires an adequate representation of human-modified landscapes and an explanation of the relationships between socioeconomic and biophysical factors (Etter et al. 2011), which is facilitated through the development of BNs and influence diagrams.
For example, Stelzenmüller et al. (2010) developed a BN-GIS linked framework to visualize relationships between cumulative human pressures, sensitive marine landscapes, and landscape vulnerability in an effort to assess the consequences of potential marine planning objectives, and to map uncertainty-related changes in management measures. The marine landscape provides ecosystem services, which, in this case, all relate to natural resource use, such as fishing and oil and gas extraction. However, those services are based on human activities that exert pressure on the marine system and make it vulnerable to the extent that those services might no longer be provided. Based on data from the literature, the authors developed relationships between changes in human pressures such as aggregate extraction, fishing, and development of oil and gas infrastructure on the vulnerability of the marine landscape.
Another example is provided by Smith et al. (2007), who developed a BN-GIS linked framework to explore habitat suitability for the Julia Creek dunnart (Sminthopsis douglasi), an endangered ground-dwelling mammal of the Mitchell grasslands of north-west Queensland, Australia. Expert knowledge, supported by field data, was used to determine the probabilistic influence of grazing pressure, density of an invasive shrub, land tenure, soil variability, and seasonal variability on dunnart habitat suitability. They carried out a sensitivity analysis to identify the influence of environmental conditions and management options on habitat suitability. Through the model, the authors were able to identify that management efforts focused on maintaining areas of low grazing pressure and low prickly acacia density on clay soils demonstrated the highest probability of providing enough suitable habitat to conserve dunnart populations. Given the significance of habitat to ecosystem services (de Groot et al. 2002), this example demonstrate the use of BNs and influence diagrams in structuring complex information and relationships to support decision making.
The ability to organize and make knowledge operational is a key part of understanding the use of BNs and influence diagrams (Haines-Young 2011). Because influence diagrams can be used to represent the current understanding of causal relationships between variables, these diagrams can be used as a focus of discussion between experts and stakeholders to elicit differing views about key elements in a system and how the system might work. From a technical perspective, BNs and influence diagrams operationalize system knowledge by expressing the strength and certainty about system relationships as a set of conditional probabilities (Haines-Young 2011). When linked to GIS, these tools provide a powerful means for eliciting how stakeholders view and understand a system in concrete land-use terms, and ultimately provide a more robust systems-based understanding across all participants to the decision-making process (Johnson et al. 2012).
MCDA methods provide another integrating framework for the results of disparate analyses and stakeholder preferences (Kiker et al. 2005). Again, the link to GIS-based data and model results enhances the use of MCDA with respect to ecosystem service endpoints, and there are some examples of linked GIS-MCDA in the literature (for a review through 2004, see Malczewski 2006), particularly in the context of land-use decisions. Linking GIS directly to the MCDA allows flexibility in how the model is specified. Alternatives assessment in MCDA requires the development of criteria against which the alternatives are measured, and the regulatory context of the MCDA application can guide that process (see Kiker et al.  for an example of an MCDA application specifically within a Superfund risk assessment). Criteria can be developed for ecosystem services directly or through the use of proxies (e.g., habitat quality). Setting up and specifying the MCDA model is an iterative process requiring the input of stakeholders, decision makers, and analysts, and needs to incorporate the particulars of the regulatory framework (e.g., risk assessment), if applicable.
Alternatively, ecosystem services can be expressed through the overall objectives for the alternatives. For example, stakeholders or decision makers define the outcomes that are important to them, such as ability to catch and eat fish, or habitat protection for a particular species, or flood control. Knowing those are the attributes to be maximized, the analyst then designs the criteria for the MCDA accordingly.
A remediation example
Evaluating alternatives for potential remediation of a contaminated site provides an example context for demonstrating the use of decision analytic approaches. The site consists of a Pb-contaminated upland area, with an adjoining waterfront and aquatic area contaminated primarily by PCBs as shown in Figure 1. Both portions of the site represent important recreational areas for the community, and several proposals exist for use of the upland site, largely focused on developing a playground with shared urban greenspace for recreational waterfront use. The ability to fish along the pier is another goal, and although catch and release fishing is permitted, fish consumption advisories are in place due to the monitored PCB levels in caught species. Other community members place a greater emphasis on providing suitable habitat for wildlife. Eliciting land-use preferences across stakeholders leads to the identification of both potential alternatives as well as specific criteria against which to evaluate those alternatives in the MCDA model.
Sustainability in the context of remediation implies different outcomes to different people, highlighting the use of a decision analytic approach in eliciting priorities, both in terms of strategic outcomes and metrics—in this case, stakeholders agreed on a set of criteria that did not include such aspects as climate or energy footprint of the alternatives under consideration (Butler et al. 2011). As decision making at the site largely followed a Superfund approach, the emphasis was on potential human health and ecological risks and the services that remediation of the site would provide (e.g., fishing, recreational space, and so on). Clearly different outcomes and metrics are possible, and can be determined through a collaborative process across diverse stakeholders, given a particular regulatory framework. Decision analytic approaches facilitate that communication.
The ecosystem services provided at this urban location were primarily focused on recreational activities for the community, ranging from shared greenspace with and without playgrounds to fishing along the pier. In fact, these represent the strategic outcomes that the community would like to achieve, and implicit in those outcomes are the conditions under which they can occur, which define the specific alternatives to be considered. That is, concentrations of Pb in soil will need to be addressed to use the greenspace as a playground. The ability to consume fish that are caught requires that PCB concentrations be decreased in sediment. Consequently, the following remedial alternatives were proposed for the area:
A0: Take no action, i.e., maintain the status quo
A1: Remediate the upland soils only
A2: Remediate sediments only
A3: Remediate both the upland soils and the sediments
A4: Develop and implement institutional controls in place of remediation
Criteria against which to evaluate the effectiveness of the alternatives were developed collaboratively by identifying specific endpoints and outcomes, including:
Blood–Pb concentrations in children and adults (as a measure of human health risk) estimated from a GIS data layer together with the EPA IEUBK model
Incremental lifetime cancer risk to humans (from ingestion of PCB-contaminated fish estimated using the FishRand bioaccumulation model)
Noncancer health effect hazard indices (HIs) (Child HI and Adult HI) for humans potentially exposed to PCBs (from ingestion of PCB-contaminated fish estimated using the FishRand bioaccumulation model)
Ecological risk, population-level hazard indices for ecological receptors from Pb and PCBs (estimated from the SEEM model for upland receptors and FishRand for fish-eating birds and mammals)
Habitat, habitat suitability index for the upland portion of the site; combined with a data layer from sediment profile imaging data for the aquatic portion of the site
Cost associated with the various remedial alternatives
Community acceptability of each alternative derived from a participatory stakeholder process
Implementation of the alternatives will result in changes to contaminant concentrations in the environment that can be visualized based on the underlying databases. These changes are fed into models used to predict the raw values for each of the criteria, which will necessarily differ across alternatives and will be normalized within the MCDA model (Kiker et al. 2005) to facilitate integration across disparate endpoints. For example, the FishRand bioaccumulation model, a spatially explicit, probabilistic model, was used to predict population distributions (with associated uncertainty) in PCB concentrations to which humans and ecological receptors are exposed (von Stackelberg et al. 2002). These predicted concentrations are in turn used to develop human health and ecological risk assessments. In this case, best estimates taken from full distributions for predicted incremental lifetime cancer risks ranged from 10 to 200 in 1 million across the alternatives. Once fully incorporated into the decision-analytic model, the result is a ranking of alternatives based on the predicted criteria values and weights. Note that the actual ranking of alternatives can be less important than the sensitivity of the rankings to the underlying assumptions, which provides important information for decision makers.
Each of the criteria will have different weights depending on the individual preferences across stakeholders and decision-makers. This is one of the most critical aspects of the decision analytic process and has been the focus of much attention in the literature (Martunnen and Hämäläinen 1995; Linkov et al. 2006; Hämäläinen and Alaja 2008; Glass et al. 2013), which will not be addressed here. However, this framework offers the opportunity to explore the sensitivity of the rankings across alternatives to the criteria weights. The ability to explore the sensitivity of the results to underlying values and assumptions is in many ways more important with respect to informing decision making than the actual ranking of alternatives, which can seem prescriptive to some decision makers. If a particular ranking is very sensitive to the weights assigned to the criteria, then that is important information for the decision maker, and provides justification for verifying those weights.
For example, the right-hand side of Figure 2 shows the impact of the weighting factor assigned to blood Pb level and how sensitive the ranking of alternatives is to changes in that weight. Currently, the weighting assigned to the blood Pb criterion is 0.146, which leads to the ranking of alternatives A3 > A4 > A2 > A1 > A0. The bottom portion of the right-hand side of the figure allows users to move the weighting (either less or more) and shows how the rankings change as a function of changes in the weighting assigned to blood Pb level. At approximately 0.43, the rankings shift and A4 becomes the preferred alternative. In other words, stakeholders would need to double the weight assigned to blood Pb for A1 to become the most highly ranked alternative. This is important information, particularly if there is disagreement or uncertainty regarding particular weights and in cases for which the cut-points are closer together.
Another example of a sensitivity analysis is shown in the left-hand portion of Figure 2. The top left portion of Figure 2 shows the relationship between the range in original units of predicted blood Pb levels across alternatives (x-axis) and the normalized values (0–1 on the y-axis) used in the model. In general, the default assumption is a linear relationship; however, this relationship can take different forms depending on the parameter, and the red square in the top portion of the figure is a toggle that allows users to change that relationship (e.g., linear, exponential, and so on). The bottom portion of the graph then shows how the ranking of alternatives changes, or how sensitive the ranking is to the form of the normalization. In this case, changing from a linear to an exponential relationship for blood Pb level leads to a reversal of alternatives A2 and A4, but the ranking for the first 3 does not change, indicating that the results are insensitive to the normalization method.
Related to sensitivity is uncertainty, and it is important to note that most commercially available decision analytic software, or programming platforms (e.g., Analytica) will provide the ability to represent uncertainty across both criteria values and weights. Although all analyses can be deterministic, in general, probabilistic frameworks provide greater information for the decision maker and perspective on the range of possible results (Thompson and Graham 1996), even if the results ultimately are collapsed to a single value.
Figure 3 provides sample output for the influence of uncertainty on the ranking of alternatives for the remediation example. The left-hand side of the figure uses interval mathematics for the criteria weights and normalized criteria values. That is, each input was defined by 4 values (a lowest possible, a lowest probable, a highest probable, and a highest possible) rather than a single value. In this case, the likeliest values for the preferred alternative (A3) do not overlap with other alternatives, suggesting that the result is robust. However, for the second (A4) and third (A2) alternatives, the left-hand side of Figure 3 shows that the probable ranges overlap, in other words, it is difficult to distinguish between these alternatives. The right-hand side of Figure 3 shows the results collapsed to single values; again, the distinction between A4 and A2 is slight, although the preferred alternative (A3) is clearly the highest. This type of analysis becomes even more important in cases where the top-ranked alternative overlaps with other alternatives such that it is difficult to distinguish which one optimizes the criteria as they have been defined. That would suggest further analysis, either in terms of more precisely defining criteria values, or perhaps adding in additional criteria that would assist in further distinguishing across alternatives.