Effect of expert opinion on the predictive ability of environmental models of bird distribution

The construction of predictive models of species distribution for conservation and regional planning can be facilitated by automatic procedures employed for the selection and transformation of predictors. It has been claimed, however, that empirical predictive models benefit from the inclusion of expert opinion at different stages of the model-building process. This is a time-consuming task that is limited by the availability of experts and difficult to standardize. Automated procedures for predictor selection based on statistical criteria are faster and easier to integrate into a geographic information system and may render highly explanatory models that fit the data with which they were built. But these models do not necessarily predict independent observations well and cannot be used to extrapolate to other areas. On the contrary, supervised models may include more frequently causal relationships, and therefore may more accurately predict new observations and extrapolate better to other areas. We built predictive models for the presence/absence of 10 bird species in two areas of Andalusia (southwestern Spain) to compare three different procedures for predictor selection ranging from a completely unsupervised to a fully supervised method. We evaluated models in three ways: (1) with the same data used to build the models, (2) with a different evaluation data set, and (3) with data from a different geographic area. The increase in the degree of expert input during model construction resulted in a significant decrease of model predictive ability when evaluated with an independent data set, and did not improve the predictive ability of the model when transferred to a new area. Unsupervised models had a greater tendency to overfit the building data, but this did not negatively affect model predictive ability or transferability to a new area. Incorporating expert opinion in the model-building process neither rendered better models as measured by their predictive ability nor resulted in models that were better suited to other regions. Therefore, unsupervised fitting procedures seem to be an adequate and cost-effective way to proceed when the aim is to generate potential distribution maps of species in a regional context.