Continuous predictors of species distributions support categorically stronger inference than ordinal and nominal classes: an example with urban bats

Understanding of how species distributions are driven by landscape-level processes has been obscured by null or inconsistent findings from poorly designed studies. We explore how differences in the way potential drivers of species distributions are defined can influence their perceived effects. Specifically, we evaluate how much statistical power is lost when continuous variables are discretised, and how the use of qualitatively defined nominal variables impacts biological interpretation of results. We fitted generalized linear models to dependent variables relating to bat distribution (species richness, diversity, relative abundance of functional groups and individual species) obtained from 36 sites across Melbourne, Australia, and independent variables that were continuous (percentage tree cover, dwelling density), ordinal (dichotomised continuous variables) or nominal (land-use, urban context). We found that models fitted with continuous predictors had better fit and explanatory power than those fitted with ordinal predictors for all response variables. Ordinal models failed to detect statistically significant effects for 4 of the 11 response variables that were successfully modelled with continuous data, suggesting Type II errors had occurred. Models fitted with nominal data explained a comparable amount of variation in some dependent variables as continuous models. However, interpretation of the mechanisms behind responses to nominal categorical levels was obscured because environmental conditions within them were confounded and not homogenous. To gain better understanding from nominal predictors would therefore require further investigation. Our findings show that careful consideration must be given to the choice of environmental variables used for species distribution modelling and how those variables are defined.