Bayesian Belief Networks as a tool for evidence-based conservation management

Summary

Effective conservation management is dependent on accessing and integrating different forms of evidence regarding the potential impacts of management interventions. Here, we explore the application of Bayesian Belief Networks (BBN), which are graphical models that incorporate probabilistic relationships among variables of interest, to evidence-based conservation management. We consider four case studies, namely: (i) impacts of deer grazing on saltmarsh vegetation; (ii) impacts of burning on upland bog vegetation; (iii) control of the invasive exotic plant Rhododendron ponticum; and (iv) management of lowland heathland by burning. Each of these themes is currently a significant conservation issue in the UK, and yet the potential outcomes of management interventions are poorly understood. Through these examples, we demonstrate that BBNs can be used to integrate and explore evidence from a variety of sources, including expert opinion and quantitative results from research investigations. Incorporation of such information in BBNs enables different sources of evidence to be compared, the potential impacts of management interventions to be explored and management trade-offs to be identified. BBNs also offer a highly visual tool for communicating the uncertainty associated with potential management outcomes to conservation practitioners, and they can also be readily updated as new evidence becomes available. Based on these features, we suggest that BBNs have outstanding potential for supporting evidence-based approaches to conservation management.

Introduction

Recent years have witnessed growing interest in evidence-based approaches to conservation, reflecting widespread concern that much conservation practice is based on tradition or the experience of practitioners, rather than on the results of scientific research (Pullin & Knight (2001), Pullin & Knight (2003); Pullin, Knight, Stone, & Charman, 2004; Sutherland, Pullin, Dolman, & Knight, 2004). This process has been inspired by the ‘effectiveness revolution’ that has occurred in medicine during the past 20 years, aimed at incorporating the results of medical research into medical practice (Egger, Smith, & Altman, 2003; Stevens & Milne, 1997). Evidence-based frameworks have subsequently developed in other areas of public policy, including psychology (Petticrew, 2001), education (Nye, Schwartz, & Turner, 2005), social welfare (Stagner, Ehrle, & Reardon-Anderson, 2003), and criminology (Gadon, Cooke, & Johnstone, 2005).

The common element of these evidence-based frameworks is the process of ‘systematic review’. All reviews are retrospective, observational research studies and are therefore subject to systematic and random error (Cook, Mulrow, Haynes, & Brian, 1997). The quality of a review therefore depends on the extent to which scientific methods have been used to minimise error and bias. Systematic reviews locate data from published and unpublished sources, critically appraise methods using pre-defined criteria and synthesise evidence to provide empirical answers to research questions. They differ from conventional reviews in that they follow a strict methodological and statistical protocol making them more comprehensive, minimising the chance of bias and improving transparency, repeatability, and reliability. Rather than reflecting the views of authors or being based on a (possibly biased) restricted sample of literature, they provide a comprehensive assessment and summary of available evidence (Cook et al., 1997).

Guidelines for the production of ecological systematic reviews have been established (Pullin & Stewart, 2006) and the results of the first systematic reviews of conservation evidence are now becoming available (Stewart, Coles, & Pullin, 2005; Tyler, Pullin, & Stewart, 2006). Further unpublished and ongoing reviews can be obtained from http://www.cebc.bham.ac.uk/. These reviews have focused on the identification of experimental and monitoring evidence most analogous to the controlled trials, observational studies, and diagnostic test trials synthesised in medical meta-analyses (Egger et al., 2003). Criteria for inclusion in such meta-analyses include the presence of a quantitative comparator before and after application of an experimental treatment, and/or between experimental treatment and controls.

These initial systematic reviews of conservation evidence have highlighted the fact that experimental investigations meeting these criteria are relatively rare, even for management approaches that are widely used. For example, of 317 articles with relevant titles concerning the impact of burning on blanket bog, only eight (2.5%) had comparators allowing quantitative synthesis (Stewart, Coles, & Pullin, 2004). Similarly, reviews regarding burning of dry heathland, the impact of wind farms on bird abundance, and bracken control utilised 1.7%, 12%, and 4.2% of material with relevant titles respectively (www.cebc.bham.ac.uk). As a consequence, the results of meta-analyses can lack statistical power (as a consequence of small sample sizes), especially when numerous effect modifiers are included in the analysis. Despite the rigour underlying the methodology, review outcomes are therefore often highly tentative, allowing few firm conclusions to be drawn. Such reviews can clearly be used to identify experimental knowledge-gaps, allowing priority areas for needs-led research to be identified. However, additional forms of evidence or information exist, and can be retrieved using systematic review methods (for example, published studies that fail to meet the criteria for inclusion in a meta-analysis, perhaps because a suitable comparator or control was not included). The inclusion of such information could increase the utility of reviews to practitioners, allowing tentative management recommendations to be made on all the available evidence rather than just a subset.

Methods are therefore required that would enable additional forms of evidence to be incorporated into analyses. Although experimental investigations of conservation management interventions are relatively few, much information is collected during environmental monitoring activities. Conservation practitioners often also possess deep working knowledge of the ecological communities with which they are familiar, based on the results of their practical experience and anecdotal observations of the outcome of management interventions. An analytical approach is needed that would enable such types of information to be analysed together with more rigorous scientific evidence, in an integrated manner. Ideally, the outcomes of such analyses should be made available in a form that can readily support decision-making, including the development and implementation of appropriate conservation policies.

We propose that Bayesian Belief Networks (BBN) offer a uniquely powerful tool to address these problems, by providing a structured combination of diverse lines of evidence. BBNs have developed at the interface between statistics, applied artificial intelligence, and expert system development (Pearl (1986), Pearl (1988)). BBNs are graphical models that encode probabilistic relationships among variables of interest (Bøttcher & Dethlefsen, 2003; Heckerman, 1996), and may be considered as tools for graphically representing the relationships among a set of variables (Castillo, Gutierrez, & Hadi, 1997). A BBN comprises a network of nodes connected by directed links, with a probability function attached to each node (Jensen, 2001). BBNs are therefore statistical models of a domain. The network of a BBN is referred to as a directed acyclic graph (DAG), which is used to model a domain containing uncertainty, and therefore provides a tool for reasoning under uncertainty (Jensen, 2001). This uncertainty can arise owing to an imperfect understanding of the domain, incomplete knowledge of the state of the domain, randomness in the mechanisms governing the behaviour of the domain, or any combination of these.

Bayesian networks evolved in the early 1990s, based on a deep body of theory developed for graphical models in general. Statistical graphical models have a history that can be traced back to Wright (1934) who developed them in the context of path analysis. In order to develop the directed graphs used in BBNs, several challenges had to be overcome, as described by Spiegelhalter, Dawid, Lauritzen, and Cowell (1993). The calculus of probability underlying Bayesian networks was at one time considered to be both epistemologically inadequate and computationally infeasible for complex domains. The principal difficulty was that complex applications require the specification of what can be huge joint probability distributions. In addition, evidence propagation within such a framework requires the computation of probabilities for events of interest that are conditional on what could be arbitrary configurations of other variables. The computational hurdles have now been overcome, due in large part to the advances derived from the seminal work of Pearl (1986), Pearl (1988). Later, Pearl (1995) showed how graphical models can be used for causal inference, thus strengthening the underlying justification behind contemporary applications of BBNs that use expert knowledge to determine both the structure and parameters of the networks (Spiegelhalter & Cowell, 1993). As a result of software developments and the increased availability of computing power, construction of large BBNs is now feasible (Neil, Fenton, & Nielson, 2000), and the method can readily be implemented on a personal computer.

Contemporary software programs for implementing BBNs are extremely flexible. BBNs can be built directly from knowledge of the domain of interest. Alternatively, it is now possible for both the structure and the parameters of a BBN to be learnt directly from a data set, and for this reason, the method is widely used for automated data mining, particularly for market research. However, many of the issues identified by conservation managers (see Sutherland et al., 2006) do not generate the large quantities of replicated data needed for such data mining. In such circumstances, it is best to view BBNs as decision-support tools helpful for combining expert knowledge with available empirical data (Marcot, Holthausen, Raphael, Rowland, & Wisdom, 2001).

Bayesian analytical techniques have received growing interest from ecological researchers since publication of a special edition of Ecological Applications in 1996 (Crome, Thomas, & Moore, 1996; Ellison, 1996; Gertner & Zhu, 1996). Typically, Bayesian statistics are used to find parameter values when the stochastic component of a model is represented by one or more continuous probability density functions. The directed acyclical graphs used to represent these models can follow the same formalisms as BBNs. However, examples of the use of BBNs in ecology or resource management are few. Examples include predicting density of mountain aspen suckers (Haas, 1991), assessing population trends in aquatic and terrestrial vertebrate species (Marcot et al., 2001; Rieman et al., 2001), integrated water resource planning (Bromley, Jackson, Clymer, Giacomello, & Jensen, 2005), social aspects of resource management (Cain, Batchelor, & Waughray, 1999) and assessing the impact of commercialising non-timber forest products on livelihoods (Newton et al., 2006). We are not aware of any previous attempt to explore the potential value of BBNs specifically to evidence-based conservation management.

In this paper, we explore the application of BBNs to evidence-based conservation management through consideration of four case studies. These illustrate a range of different conservation issues and management options, and vary in the type and quality of evidence available. In two of the examples, BBNs are constructed that incorporate the results of systematic surveys of the conservation literature, and in one case, this form of evidence is contrasted with that based on the experience of conservation practitioners. In each case, BBNs are used to assess the potential impact of management interventions on some outcome or variable of conservation interest. The four examples are: (i) impacts of deer grazing on saltmarsh vegetation; (ii) impacts of burning on upland bog vegetation; (iii) control of Rhododendron ponticum; and (iv) management of lowland heathland by burning. Each of these themes is currently a significant conservation issue in the UK, and yet the potential outcome of management interventions is uncertain.