How believable is your BBN

Bayesian Belief Networks or BBNs are gaining prominence in ecology. They are a powerful and attractive tool for managing and understanding complex processes because they represent the process graphically, where each node in the network represents either the prior or conditional probability of the parameter of interest. Causal links are represented by arcs that join to nodes and indicate the dependencies between nodes in the network. The fundamental idea behind these causal relationships is Bayes’ theorem, which provides a premise for combining the prior and conditional probabilities assigned to each node of the BBN to form posterior estimates of the quantities of interest that can be readily updated. Despite their popularity and wide spread use in solving complex, large scale ecological problems of importance there are a number of issues relating to the structure of the model and incorporation of expert information that modellers need to be aware of. This leads us to ask “How can we really believe the output from a BBN, given that a large proportion of the information feeding into the network often relies on expert opinion that may be inaccurate or carry hidden biases?” Analysts are faced with three important tasks when producing a BBN: (1) identifying the variables pertinent to the problem at hand; (2) identifying the relationships between these variables; and (3) expressing these relationships as a series of conditional probabilities. Through this implementation a number of issues can arise resulting in an ill-defined BBN which inadequately reflects the underlying system processes nor captures and reflects the expert opinion defining these processes well. Using a popular BBN software package, Netica™, we examine some of the potential pitfalls of BBNs through two specific examples. The first focuses on a fishery management problem, which examines whether management should remain passive or seek a more active approach to commercial fishing. The second investigates the quarantine risks associated with the importation of commodities and examines the probability that pests or diseases will enter a nation with imports of goods. We examine issues of discretisation (process of converting a continuous probability distribution to one that is discrete), scaling (process of transforming data), complexity (number of nodes and linkages) and network structure (nodes and linkages defining causal relationships) in the context of these two examples and show that each can have a dramatic impact on the posterior probabilities of each model investigated and therefore have the potential to impact management decisions if implemented. In our analysis, we show that although BBNs offer a powerful mechanism for capturing both expert information and empirical data (where available) and can also address issues of uncertainty through the elicitation of conditional probabilities, there are some aspects of BBNs that the modeller needs to consider. We highlight these issues and suggest ways in which to guard against these problems so that BBNs are used more appropriately for the modelling exercise being considered.