Integrating Expert Perceptions into Food Web Conservation and Management

Decision‐makers often rely on expert knowledge, especially in complex and data‐poor social‐ecological systems (SESs). However, expert knowledge and perceptions of SES structure and function vary; therefore, understanding how these perceptions differ is critical to building knowledge and developing sustainability solutions. Here, we quantify how scientific, local, and traditional knowledge experts vary in their perceptions of food webs centered on Pacific herring—a valuable ecological, economic, and cultural resource in Haida Gwaii, BC, Canada. Expert perceptions of the herring food web varied markedly in structure, and a simulated herring recovery with each of these unique mental models demonstrated wide variability in the perceived importance of herring to the surrounding food web. Using this general approach to determine the logical consequences of expert perceptions of SES structure in the context of potential future management actions, decision‐makers can work explicitly toward filling knowledge gaps while embracing a diversity of perspectives.


Introduction
Experts play a key role in decision-making in conservation. In the absence of certainty about the nature and behavior of complex social-ecological systems (SESs), expert opinions are often elicited in hopes of separating matters of fact from matters of value to complement existing data and inform conservation decisions (Dietz 2013). Typically, technical experts communicate their understanding of social-ecological processes to decisionmakers, enabling them to rely upon best available science (Ryder et al. 2010). For example, assessments of oil spill impacts (e.g., major oil spills; Leschine et al. 2015), climate change mechanisms (e.g., IPCC 2014), and poten-tial tradeoffs associated with scientific whaling (e.g., the case for scientific whaling; de la Mare et al. 2014) have all relied on expert judgment (reviewed in Redpath et al. 2013). However, experts can exhibit high levels of uncertainty because knowledge integration among individuals is inherently complex (Raymond et al. 2010;Drescher et al. 2013), and in some cases not possible (Gray et al. 2012).
Despite the potential for each expert's training, experience, and education to guide judgments (Burgman, Carr et al. 2011), many conservation decision-making processes focus on gathering input from select individuals with substantial, but not necessarily objective, information about a given topic (Burgman, Carr et al. 2011;Martin et al. 2012;Drescher et al. 2013). Indeed, divergent views are common for two reasons. First, expert perceptions of SESs are typically based on what they have learned from experience or in the classroom (i.e., human cognition). Second, the terms employed to describe SESs (e.g., ecosystem structure) are human constructs, and the way in which perceived differences are discussed is typically qualitative, imprecise, and prone to biases in human reasoning (but see Doswald et al. 2007). For instance, differences in how the species concept is viewed by scientists (Levin 1979) have led to disputes (Hey et al. 2003), with important implications for conservation of imperiled species (e.g., Waples 1991;Beaudreau et al. 2011).
Here, we document varying expert perceptions of ecosystem interactions in the Northeast Pacific Ocean and explore their implications for conservation and management. In this region, Pacific herring (Clupea pallasii) are a key ecological, economic, and cultural resource. Numerous terrestrial and marine coastal organisms, including several commercially harvested fishes, prey on herring (Schweigert et al. 2010). Herring were a major focus of industrial seine and gillnet fisheries, but stock collapses in the 1960s and 1990s resulted in fisheries closures (DFO 2014) and conservation concern. Herring roe is also an important cultural and subsistence resource for a number of First Nations (Jones et al. 2010). Because of the central role of herring in Northeast Pacific SESs, fisheries closures and subsequent reopenings have led to tensions surrounding herring and herring fishing among First Nations groups, the Canadian government, and commercial fishery interests (DFO 2014).
Given the numerous social and ecological connections to herring, we explored how experts from a variety of backgrounds perceived Northeast Pacific ecosystem interactions. We asked each expert to describe the number, direction, and strength of food web interactions among functional groups connected directly or indirectly to herring. Based on these responses, we constructed Fuzzy Cognitive Maps (hereafter, cognitive maps) of the herring ecosystem, revealing each expert's unique perception of the number, strength, and direction of relationships among network nodes, and how perceptions of different experts related to one another. We also simulated responses of the cognitive maps to hypothetical scenarios, including an increase in herring, the continued recovery of humpback whales (a key herring predator), and changes in ocean productivity, asking if the logical consequences of differences in perception of ecosystem structure magnify or diminish variability in predictions about ecosystem responses to management actions.

Fuzzy cognitive maps
To quantify variation in experts' perceptions of the structure and function of the herring-centric food web in the Northeast Pacific Ocean, we collected cognitive maps from a range of scientific experts. Cognitive maps are basic mathematical and graphical representations of an individual's perception of the number and strength of relationships among nodes in a network. In this case study, the network is the Northeast Pacific Ocean food web, and the nodes in the network are the functional groups. Knowledge constructed in this manner can externally represent an individual's organized understanding of the workings of the world around them (Gray et al. 2014). These representations of understanding can then be manipulated mathematically to indicate the logical consequences of an individual's perceptions based on their understanding of the dynamics of the external world. For example, by increasing or decreasing key variables as continually high or low (referred to as "clamping"), future scenarios, such as the increased abundance of a predator, can be simulated given a specific set of perceived linkages and interaction strengths (Özesmi &Özesmi 2004). This clamping is conducted until the system reaches a new equilibrium that can be compared to the steady state-the equilibrium relative abundance in the absence of a perturbation (for additional detail, see Supplementary Material).

Expert elicitation
To build cognitive maps of the herring ecosystem in the Northeast Pacific Ocean, we identified 14 key functional groups in the herring food web (Table S2), based on published literature (Ainsworth et al. 2008;Schweigert et al. 2010;DFO 2014), the authors' natural history knowledge of important ecosystem interactions, and pilot testing with five experts inside and outside of our focal area in Haida Gwaii. While providing a particular set of functional groups can constrain cognitive maps of the system, it allows for quantitative comparisons among experts (Gray et al. 2014). Experts were defined as having scientific (e.g., agency or university scientists), local (e.g., long-term residents), or traditional (i.e., First Nations) ecological knowledge and/or practical experience in the Northeast Pacific Ocean herring ecosystem. Experts were identified through stratified chain referral sampling (Biernacki & Waldorf 1981), which yielded a complete sample of 27 experts. We then explored the potential role of training, experience, and cognition as key factors that may influence the diversity of expert perceptions (following Burgman, Carr et al. 2011;Morgan 2014).
We asked each expert their perception of the number and strength of interactions between all pairs of functional groups. Respondents were also given an opportunity to provide information on their uncertainty about interactions (Table S2). Interaction strength elicitations ranged from -2 (strongly negative) to +2 (strongly positive), and were scaled from -1 to +1 for analysis (for additional detail, see Supplementary Material). We also asked a series of demographic questions detailing information that could potentially influence responses (e.g., age, years of experience, professional affiliation, and place of residence; Table S1).

Network analysis of cognitive maps
We conducted a network analysis to describe the geometry and strength of interactions for each cognitive map and then subjected the resultant metrics to hierarchical cluster and nonhierarchical partitioning analyses. The network analysis metrics we used to represent herring ecosystem structure included number of connections in each food web, average of the absolute value of the interaction strengths, centrality of four key functional groups of conservation interest, a hierarchy index, and number of transmitters, receivers, and ordinary concepts suggested byÖzesmi &Özesmi (2004; Table 1).

Analysis of expert perceptions of food web structure
Demographic characteristics were not effective predictors of variation in perceived ecosystem structure (Table S3). We therefore applied nonhierarchical partitioning analysis to ask whether evidence existed for ࣙ2 clusters of experts based on the similarity of cognitive maps, summarized in terms of the network metrics. To visualize the distances between experts, we calculated Euclidian distances between each pairwise combination of experts based on the network metrics described above, and used hierarchical cluster analysis to identify potential groupings of experts (using an agglomerative average linkage method; Venables & Ripley 2002). Hierarchical cluster analysis makes no assumptions about a priori relationships among experts (e.g., demographic characteristics) but rather searches for a posteriori groups based on the differences among individuals in cognitive map structural metrics, allowing comparison of expert knowledge by the nature of their understanding as opposed to membership in a demographic group (for additional detail, see Supplementary Material).

Scenario analysis: perturbing the herring food web
There is a fair amount of uncertainty surrounding the future of Pacific herring in the Northeast Pacific. This uncertainty is rooted in the complex social and ecological influences on the species, all of which occur at a range of spatial scales. We evaluated the functional consequences of each expert's perceived ecosystem structure by simulating three perturbations, each of which caused a consistent increase (press perturbation) in a single node in the food web until all nodes in the food web reached a new equilibrium. Specifically, we Specifically, we simulated the following: (1) an increase in humpback whales concordant with projected humpback population growth (Ford 2009), (2) an increase in herring-a simulation in accordance with the desired trajectory of the depleted stock (DFO 2014), and (3) an increase in zooplankton-analogous to a regime of cold, nutrient-rich water years that support productive zooplankton populations (Mantua & Hare 2002; Figure 1). We conducted scenario analyses on each cognitive map (n = 27) and measured the change in relative abundance of each of the 14 functional groups compared to its relative abundance at equilibrium in the absence of a perturbation. Such an approach is expected to represent predicted outcomes under different ecological change scenarios across different types of experts.

Statistical analysis of expert perceptions of ecosystem function
As with the analysis of expert perceptions of herring ecosystem structure, we used hierarchical cluster and nonhierarchical partitioning analyses to ask whether subsets of experts perceived ecosystem responses similarly. Importantly, these predictions represent the logical consequences of information elicited from experts, rather than direct elicitations from scenario-based questions. Though we tested whether expert demographics were effective predictors of variation in responses to perturbations from the three scenarios, we found none (Table S2). Thus, in this application, we sought clusters of experts based on the expected percent change in relative abundance of each of the 14 functional groups under each scenario. We also estimated two ecosystem responses for each of the three scenarios: (1) average percent change in abundance of all 14 functional groups and (2) ecosystem reorganization index, which estimates discordance among functional groups in their response to a scenario, relative to the aggregate response of all functional groups (sensu Samhouri et al. 2010). Components which only have "forcing" functions; indicates number of components that affect other system components but are not affected by others N (receiver) Components which have only receiving functions; indicates the number of components that are affected by other system components but have no effect N (ordinary) Components with both transmitting and receiving functions; indicates the number of concepts that influence and are influenced by other concepts Centrality Absolute value of either (a) overall influence in the model (all + and − relationships indicated, for entire model) or (b) influence of individual concepts as indicated by positive (+) or negative (−) values placed on connections between components; indicates (a) the total influence (positive and negative) in the system or (b) the conceptual weight/importance of individual concepts (Kosko 1986a

Contextualizing our approach within existing mental model approaches
Our approach advances existing methods focused on building cognitive maps to improve conservation and management (Biggs et al. 2011). Researchers use several methods to collect and evaluate mental models and shared beliefs in natural resource management (Lynam & Brown 2011). For example, some methods assume homogeneity in knowledge within demographic groups, despite examples of ample heterogeneity within knowledge groups (e.g., Gray et al. 2012) and at various scales  (Table 2 ), which are plotted univariately for each expert group in boxplots (c-g). In box and whisker plots, the upper and lower "hinges" correspond to the first and third quartiles (the 25th and 75th percentiles) and whiskers represent 1.5 times the interquartile range.
exploring variation in individual cognitions assumes perceptions are linked to demographic backgrounds a priori, whereas here we test for links between demographic backgrounds and perceptions, but also use multivariate clustering analysis to define cluster of similar experts based on perceptions of food web structure and function. Overall, our approach highlights the utility of building individual cognitive maps, which can increase stakeholder communication and facilitate the integration of multiple knowledge sources (Biggs et al. 2011), while also avoiding assumptions about links between demographic characteristics and expert perceptions.

Results
Expert experience with the herring ecosystem in the Northeast Pacific Ocean averaged 19 years (range 5-61years), yet this depth of experience did not translate into cognitive maps with highly similar network properties (Table S4). For example, networks varied widely in number of connections (range 42 to 125). Multivariate analysis suggested expert demographics did not explain variation in perceptions of herring ecosystem structure (Table S3). Instead, cluster analysis revealed two prevailing views that were unrelated to amount of experience (Figures 2a, b). An OLS regression testing for univariate relationships between years of experience and food web structural properties revealed no correlation (P > 0.15 for all structural properties). Experts with different demographic characteristics were well represented in both groups. For example, Group 1 was 50% academics, 66% NGO, 66% island residents, and 64% individuals who identified as female. Group 2 was 50% academics, 33% NGO, and 33% on island, and 64% individuals who identified as male. This high within-group variability in demographic characteristics was a major contributor to statistically nonsignificant differences based on demographic background. An additional explanation for our inability to detect statistical differences among groups is the number of experts sampled, which is somewhat low despite  Figure 3 Ecosystem response (i.e., percent change in relative abundance of functional groups) to three different scenarios simulating increases in zooplankton (a), herring (b), and humpback whales (c) averaged across all cognitive maps. Upper and lower "hinges" on box and whisker plots represent the first and third quartiles (the 25th and 75th percentiles) and whiskers represent 1.5 times the interquartile range.  Figure 4 Among-scenario comparison of ecosystem response to simulated food web perturbations. Nonmetric multidimensional scaling plot generated from change in relative abundance of functional groups in each scenario relative to a steady state (a). Each point represents an expert, with each expert represented for each of the three scenarios. Experts perceiving a similar ecosystem shift in response to each of the three scenarios are closer together. Point color represents the three different scenarios (Blue: humpback +, Green: Herring +, Gray: Zooplankton +), and point type (circle or triangle) represents the two clusters that emerged from the expert's perceptions of ecosystem structure. Univariate plot of average (±1 SE) multivariate dispersion demonstrates among expert variability in response to scenarios (b).

(b) (a)
exhaustively sampling the expert pool through stratified referral sampling. Expert perspectives of the ecosystem diverged based on several characteristics of the cognitive maps, including a 35% difference in overall influence of focal functional groups (i.e., centrality), 25% difference in interaction strengths, 15% difference in connections, and a 28% difference in whether connections were democratic (i.e., hierarchy index; Figures 2c-g).
Variable perceptions of herring ecosystem structure did not necessarily correspond to differences in expected outcomes of hypothetical scenarios, despite unique responses of each cognitive map to simulated perturbations ( Figure 3). Furthermore, backgrounds and demographic characteristics of experts did not readily explain variability in expected changes in species relative abundance (Table S4). In fact, we found cryptic agreement surrounding scenarios despite divergent perceptions of herring ecosystem structure. For example, simulated increases in herring predators led to a predicted decrease in herring, zooplankton, and other forage fishes, whereas simulated increases in zooplankton predicted increases in relative abundance of all species (Figure 4a).
The two clusters that emerged from analysis of structural metrics describing cognitive maps effectively predicted responses to hypothetical scenarios ( Figure S1). For example, food webs with more connections and higher estimated interaction strengths exhibited a greater level of ecosystem reorganization ( Figure S2). However, each of the three simulated scenarios differed in the level of among-expert disagreement, with the widest variation emerging from the simulated increase in herring (Figure 4b). Hypothetical scenarios representing expected ecosystem state related to increases in herring predator (whales) and prey (zooplankton) abundance produced variable responses among cognitive maps, but these responses did not diverge into distinct groups. In contrast, the hypothetical scenario related to an increase in herring produced two significantly divergent perspectives (Figure 5a). One expert group predicted a 182% greater reorganization of the ecosystem and a 78% higher average percent change in relative abundance relative to the other expert group (Figure 5b). These contrasting perspectives center on the strength of expected declines in relative abundance of zooplankton (typically herring prey) and other forage fish (resource and apparent competitors with herring), and expected increases in herring predators (e.g., groundfish, whales, and seabirds; Figure 5). Background and demographic characteristics did not explain the discordant perspectives between these two expert groups (Table S4). Furthermore, the demographic composition of these two groups is very similar, with 85% of the individuals from the cluster analysis of ecosystem structure found in identical groups that emerged from cluster analysis of simulated increases in herring.
For the increased predator and prey scenarios that led to relative consensus among experts (i.e., lower multivariate dispersion in community response to simulated scenarios), an understanding of differences in perception of herring ecosystem structure would have inappropriately suggested potential for divergent views over ecosystem functioning (Figure 4a). This diminished divergence of expert perceptions did not emerge from the increased herring scenario (Figure 4b). Rather, structural differences in perception of the herring ecosystem were critical predictors of functional differences where herring were the focal point of change.

Discussion
For a wide range of public policy issues, there is an increasing dependence on scientific expertise to inform decision-making (Martin et al. 2012) and a broadening expectation for experts to extend their knowledge to more disparate areas (Gibbons 1999). Many of these issues (e.g., coastal defense and Hurricane Sandy/Katrina, Ebola dynamics, and GMO foods) are directly related to how ecosystems will respond to forecasted increases in natural and anthropogenic perturbations (Turner 2010). As in other spheres, because of limited data and the urgency of decision-making, the institutional and governance structures of natural resource and conservation management increasingly rely on expert knowledge (Thuiller et al. 2008). This reliance comes despite widespread acknowledgment that expert knowledge is often incomplete, variable, and biased (Martin et al. 2012;Drescher et al. 2013). We show here that among-expert differences in perceptions of ecosystem structure are logically tied to consequences for how an individual might view the outcomes or impacts of predicted future change. Recognizing this causal chain, and quantifying it explicitly, is the first step toward navigating ecosystem-based conservation decisions that rely on expert knowledge.
Experts are susceptible to known cognitive biases due to heuristics (i.e., informal rules people use to make judgments) such as "availability," the ease with which an idea can be brought to mind, and "anchoring and adjustment," where an individual is provided a particular value or range of values and adjusts from that "anchor" (Morgan 2014). To diminish the likelihood of including these biases in our data set, we attempted to reduce variation in weighted estimates between variables and focus measurement on knowledge variation in terms of network structure, as opposed to variation in probability estimates (Morgan 2014). Through our approach, we show that among-expert differences in perceptions of ecosystem structure are logically tied to consequences for how an individual might view the outcomes or impacts of predicted future change. Recognizing and quantifying causal chains can allow experts to consider multiple factors that influence one another in a complex web of interactions, including feedbacks. The exact reasons underlying differences among expert knowledge and perceptions are unclear. Future studies would benefit from including meta-knowledge about expert knowledge, including dimensions about knowledge confidence in the relationships represented and epistemic orientations (Miller et al. 2008), to understand how different "ways of knowing" maybe more or less valued by different expert groups and influence expert knowledge representations.
Our results show that experts can exhibit divergent views about the structure of a complex ecosystem, independent of commonly identified "bins" of expertise (e.g., local, scientific, traditional). Our inability to predict variability in perceptions through demographic characteristics stands in contrast to examples from other arenas (e.g., political party affiliation and ideologies; Pinello 1999). Yet, expert backgrounds (e.g., years of experience) do not always predict expert performance (Burgman, McBride et al. 2011). Our finding reinforces the concept that expert knowledge is more fluid and pluralistic than discrete categories acknowledge (Raymond et al. 2010;Krueger et al. 2012). However, it is also possible that we did not detect links among background characteristics and perceptions because there were hidden demographic characteristics we did not test, our study was limited in statistical power, or perhaps there was some cognitive bias resulting from our elicitation method (Morgan 2014). Simulated management scenarios using cognitive maps of the herring ecosystem highlighted additional implications based on differences in perceptions of how ecosystems may respond to future perturbations. In particular, simulations of herring recovery using each expert's unique perception of food web structure demonstrated that not all experts perceive herring as having a similar number and strength of connections to the broader ecosystem and that this may lead to different predicted outcomes across the food web. These disparities in perception are particularly significant because herring sit at the center of the food web (Watts & Strogatz 1998), as is common for many marine forage species in coastal ecosystems (Cury et al. 2000). Moreover, similar variability in perception is likely to be common for complex networks with the potential to be highly centralized, dynamic, and interactive (e.g., financial systems; May 2013).
Among-expert variability in perceptions of the number and strength of connections between herring and the rest of the food web portends of variable management advice by experts when it comes to: (1) protected species (e.g., seabirds and marine mammals) that consume herring, (2) sustainable harvest of commercially valuable fishes that prey upon herring (e.g., groundfishes and salmon), and (3) marine ecosystem-based management in the Northeast Pacific. For example, experts were divided in their expectations about the impacts of a herring increase on Pink and Chum salmon: one group predicted an increase while the other predicted a decline (Figure 5b). Under the same scenario, one group of experts perceived a simulated increase in herring would lead to an 89% greater increase in whales relative to the other group (Figure 5b). These results suggest managers of the herring ecosystem are confronted with different knowledge systems and diverse perceptions that they must reconcile or reject as they weigh different (and at times divergent) forms of expertise. As in many other environmental decision-making contexts, recognition of these variable perceptions of food web structure may encourage efforts to fill knowledge gaps in areas where experts disagree. Where there is expert consensus, promoting social learning among stakeholders about commonalities in their logical chains of reasoning, despite diverse and cultural backgrounds, may be a positive force in a system where mistrust and differences in values contribute to conflict over common pool resources (Welch 2015). In contrast, mixed demographic composition within a cluster of experts with similar food web perceptions could be associated with differences in values as well as mistrust, making it difficult to find consensus (Burgman, McBride et al. 2011). By including diverse sets of expert knowledge, the total space of available knowledge increases and can be particularly useful for exploring events and processes that are outside the normal range or are difficult to test empirically. Furthermore, while variable expert perceptions can lead to conflict, it may also be a positive force through integration with adaptive management. For example, given a set of common ecosystem goals, surveys could be used to describe variation in expert perceptions and to test alternative logical chains of reasoning that compete with one another, and data which support a group of experts' knowledge can be used to validate perceptions empirically. In these cases, conflicting expert knowledge can be considered an asset, as opposed to a liability, since knowledge diversity is likely to lead to scrutiny of expert opinions, leading to more robust conservation decisionmaking.

Conclusion
Previous research has demonstrated that expert perceptions can vary widely; however, fewer studies have explored the potential implications of diverse expert perceptions for the management of complex SESs. Our findings demonstrate how the composition of expert panels will strongly influence expert perceptions of ecosystem structure, which can have cascading effects on the perceived outcomes of future management actions. As such, binning knowledge into a priori categories based on expert backgrounds can lead to erroneous conclusions; rather, embracing a diversity of knowledge in dialogue surrounding alternative management actions will help address uncertainty, can reduce conflict, and potentially improve management outcomes. Demonstrating the variety of perceptions that exist, and the potential implications of these variable perceptions given future management scenarios, is a critical step to moving forward with ecosystem-based conservation in the face of uncertainty that surrounds complex systems and their dynamics.

Supporting Information
Additional Supporting Information may be found in the online version of this article at the publisher's web site: Table S1. Number of technical experts per affiliation and gender category. Table S2. Species embedded within each of the 14 functional groups described to participants. Table S3. IPCC certainty values . Table S4. Multivariate analysis testing whether demographic characteristics predict variation in food web network metrics. Table S5. Demographic predictors of three scenarios simulating press perturbations to the food web at the bottom (zooplankton increase), middle (herring increase), and top (whale increase). Figure S1. The capacity of two clusters of experts (white and green) based on structural properties of the system to predict variation in ecosystem response to three perturbations. Figure S2. Positive correlation between a structural property of each expert's mental model of the food web and the amount the ecosystem fluctuates in response to the increased herring scenario.