Applicability of biodiversity databases to regional conservation planning in the tropics: A case study evaluation of the effect of environmental bias on the performance of predictive models of species richness

The biodiversity data typically available for fitting distributional models in the tropics come from museum and scientific collections which are often incomplete and prone to sampling and environmental biases. Nevertheless, most studies undertaken in tropical regions assume that collection data offers a satisfactory environmental coverage without any quantitative assessment. In this study, we investigate the effects of differences in environmental bias and coverage provided by distributional data when aggregated into different grid cell sizes, on the performance of species richness-environment models and predictions. We use an extensive data compilation, including national and regional collections, on the distribution of amphibians, reptiles and fishes in the hydrologic region of the Usumacinta River as a case study. General additive models and environmental variables are used to construct predictive models at 40, 20, 10 and 5 km grid resolutions, based on well-sampled cells. The best multivariate models included nonparametric interaction terms for the effects of precipitation and temperature and suggested an altitudinal shift in the relative importance of energy and water in determining the distribution of species richness. For fishes, geomorphology accounted for fine scale variation in species richness along the hydrologic network, indicated by peaks in species diversity at the junction of the major rivers where major accumulation of water and sediments occurs. For all taxonomic groups, we found that sampling biases deviated most from the mean bias at the extremes of gradients accounting for important environmental factors. The pattern of environmental bias changed with grid size, with the form and amount of change being case-specific. Biases affected distribution predictions when compared with unbiased datasets. Moreover, not all models resulted best at coarser resolution as it is commonly assumed. Our results demonstrate that bias in the available data must be evaluated before mapping biodiversity distributions, irrespective of the choice of scale.