Development of criteria for simplifying ecological risk models

Tropical seagrass habitats of the Great Barrier Reef are influenced by a complex suite of parameters. Recently there has been particular interest in the impact of land-derived runoff to the long-term viability of tropical seagrass habitats. Seagrass ecosystems are highly variable and poorly understood. A decision support tool that copes sufficiently with knowledge gaps and uncertainty was required to support the management of this complex stochastic system. We chose to develop a quantitative Bayesian Network (BN) as the risk management decision support tool for tropical seagrass. The objective of the decision support tool was to improve understanding of the current state of the seagrass ecosystem at risk, the impact of multiple threats to seagrass, and/or the impact of multiple management choices on seagrass health. A Bayesian Network was preferred for decision support tool development because BNs can summarize small-scale, unpredictable, and unmanageable processes via probabilistic expressions which can be updated as new information comes to hand. The capacity of BNs to summarize system processes encourages analytic focus to be directed toward the most critical factors. A significant issue that arises during the development of such BN applications is the large number of possible cause-effect linkages that can be modeled. Quantitative modeling of every possible threat to the ecological receptor of interest (which in this case is seagrass) will create a very complex model. This is a significant issue because complex models are difficult and time consuming to parameterize and populate. The paucity of ecological data available for parameterization further complicates modeling tasks. Expert judgment can be used to fill these gaps, but the elicitation required to access this knowledge is notoriously costly and time consuming. More complex models are also more difficult to explain and communicate to stakeholders and decision makers. This is a critical aspect of decision support tools, which are designed to be used by decision makers (i.e. managers) rather than technical experts. Inability to foster adequate understanding and acceptance of the tool among users will compromise decision support tool uptake and utilization. Clearly there is a need to simplify and focus BN modeling tasks. The best way to simplify the BN is to minimize the number of factors (representing threats) requiring parameterization. Within the context of the domain, this becomes a question of determining which system components are likely to be insignificant for achieving decision makers’ objectives. The inclusion of these low-priority factors from subsequent phases of BN modeling should be prevented. However the process used to make such modeling decisions can strongly influence the final ranking of each factor. The decision making process can thus affect the credibility of both the ranking and of the decision support tool itself. Prioritization processes for models have been developed elsewhere, however they commonly approach the task from a broad, top-down perspective directed towards ranking risk issues. We found this perspective unhelpful for prioritization of ecological factors within the issue we are modeling. Here we present a new bottom-up prioritization approach for BN model simplification. We found that a qualitative five-phased system kept the process simple, structured and focused. Mandatory documentation of evidence used (or not) to support prioritization decisions increased the rigor and consistency of the process, bolstered credibility of the outcomes among experts, and provided an audit trail. Application of the process significantly reduced the size and complexity of the seagrass conceptual model and simplified BN model construction. Comprehensive expert input was vital in the conceptual model simplification process.