Combining data-led methods with expert opinion - how Bayesian approaches can bridge the gap between academia and practice.

With hindsight, I have mistakenly avoided Bayesian approaches to ecological modelling  because they incorporate prior beliefs. The quantitative scientist in me thought that this sounded a little too vague to be of use in conservation and would surely fall foul of bias towards preconceived ideas. However, after reading the paper featured below and the detailed supplementary materials, I am now a convert to Bayesian techniques and hope to incorporate them into my work in the future.


Laws & Kesler (2012) have developed a model for selecting translocation sites for the Guam Micronesian kingfisher (GMK), Todiramphus cinnamominus cinnamominus and to me, it seems like an excellent way of combining quantitative methods with common sense whilst incorporating the complexity of issues involved in selecting suitable sites for translocation. The best way I can explain their approach is to describe their inference diagram: imagine a tree where the main thing we're interested in, island suitability, is the trunk. The trunk splits into four branches representing ecological requirements, impacts on native species, anthropogenic threats and operational support. The tree continues to branch until the generic factors associated with any translocation (e.g. presence of disease, habitat protection laws, food availabilty) give way to GMK-specific factors (e.g. West Nile virus, protected areas, insect prey). At that point, my tree analogy breaks down because some of the 'twigs' feed into several branches but hopefully, you appreciate that this is a relatively straightforward way of representing the complexity in the GMK's translocation needs.


The next job is to assign conditional probablities to each of the factors that contribute to island suitability. For habitat suitability, the two components of available area of suitable vegetation and the extent of habitat fragmentation were modelled using data from 156 Micronesian islands and the occurence of kingfishers of the same genus as GMK. This was used as training data for the GMK model to select candidate translocation sites from 239 island. The rest of the modelling process relied on qualitative decisions to set categorical outcomes, for example, if predatory non-native species were present, the island would be deemed unsuitable. These were then translated into quanitative combinations for the purposes of the judging each island's suitability (see appendix A of the paper for more details).


Only five islands were considered suitable for GMK translocation and even then, they were thought to require varying levels of management. Site visits to the five islands found two of these to be unsuitable due to degraded habitats and lack of political support. The authors caution that the models are only as good as the input data they are built on.


I can see from Laws & Kesler's paper that Bayesian methods have real potential for bridging the gap between expert knowledge and data-driven correlative methods. However, we still need people with the statistical know-how to reach across the gap. Any volunteers?

Laws, R. J., & Kesler, D. C. (2012). A Bayesian network approach for selecting translocation sites for endangered island birds. Biological Conservation, 155, 178–185. doi:10.1016/j.biocon.2012.05.016