Can we predict the presence of benthic diatom species in streams?

Accurately predicting species presence and absence is fundamental to stream bioassessment. The main goal of this research is to evaluate our ability to correctly predict site occupancy by diatom species . We examine and compare the ability of 4 different modeling techniques to predict the presence and absence of benthic diatom species. Logistic regression, hierarchical linear models, classification tree and random forest models were fit to EMAP-west data . Each model was calibrated using a subset of the data and then validated with an independent dataset. Specificity, sensitivity and kappa were used to compare the performance of each model. Preliminary results suggest that we can accurately predict the presence of common taxa; however, predictions of absence were less reliable. All models predicted occupied sites well (>96%) but had high error rates, predicting presence in as many as 67% of the empty sites. The random forest model out performed the other models by all measures, correctly predicting the presence in 97% of the sites and wrongly predicting the presence in 55%. High error rates of common taxa models and unreliable rare taxa models have important implications as predictive models become increasingly common in bioassessment and conservation biology.