Factors affecting species distribution predictions: A simulation modeling experiment

Geospatial species sample data (e.g., records with location information from natural history museums or annual surveys) are rarely collected optimally, yet are increasingly used for decisions concerning our biological heritage. Using computer simulations, we examined factors that could affect the performance of autologistic regression (ALR) models that predict species occurrence based on environmental variables and spatially correlated presence/absence data. We used a factorial experiment design to examine the effects of survey design, spatial contiguity, and species detection probability and applied the results of ten replications of each factorial combination to an ALR model. We used additional simulations to assess the effects of sample size and environmental data error on model performance. Predicted distribution maps were compared to simulated distribution maps, considered truth, and evaluated using several metrics: omission and commission error counts, residual sums of squares (RSS), and areas under receiver operating characteristic curves (AUC). Generally, model performance was better using random and stratified survey designs than when using other designs. Adaptive survey designs were an exception to this generalization under the omission error performance criterion. Surveys using rectangular quadrats, designed to emulate roadside surveys, resulted in models with better performance than those using square quadrats (using AUC, RSS, and omission error metrics) and were most similar in performance to a systematic quadrat design. Larger detection probabilities, larger sample sizes, contiguous distributions, and fewer environmental data errors generally improved model performance. Results suggest that spatially biased sample data, e.g., data collected along roads, could result in model performance near that of systematic quadrat designs even in the presence of potentially confounding factors such as contiguity of distributions, detection probability, sample size, and environmental data error.