Predicting species occurrence using information-theoretic approaches and significance testing: An example of dormouse distribution in Cumbria, UK

Predicting the occurrence of species is an essential part of conservation biology. The range of techniques used to do this has increased in recent years. This has included wider use of information-theoretic approaches; particularly Akaike’s information criteria (AIC). AIC is often used with regression modelling when predicting species distribution. The traditional method of model selection in regression modelling, stepwise significance testing, is also still widely used. This paper compares the two approaches, using the occurrence of the dormouse (Muscardinus avellanarius) in Cumbria, UK as an example. The dormouse is a protected species whose abundance and range have declined nationally. Knowledge of its occurrence in Cumbria is required in order to increase measures for its conservation in the area. The paper uses the habitat features woodland size, altitude, soil type, tree species present, temperature and rainfall as potential predictors. As only presence data was available for dormice in Cumbria, pseudo-absences were generated to allow logistic regression modelling. The use of pseudo-absences was justified using a false record permutation test. Cross-validation allowed the ability of models to predict the data to be assessed. This was judged using ROC plots. The size of the wood, temperature and whether the soil was wet or dry were the best predictors of dormouse incidence in Cumbria. The two approaches produced different models; those from the information-theoretic approach had a better ability to fit the data. The information-theoretic approach also had the advantage of enabling model averaging and provided greater understanding of the system.