Failure to account for spatial autocorrelation artificially inflates the performance of introduced species distribution models

Background/Question/Methods

Predicting the potential distribution of introduced species is frequently conducted with species distribution models (SDM). This approach quantifies the species-environment relationship in the native range and uses this association to identify areas of potentially suitable habitat in unoccupied regions.  However, these models suffer from a failure to account for the inherent spatial autocorrelation in ecological data and inaccurate quantification of uncertainty by using traditional frequentist methods of statistical inference. The potential result of this shortcoming is an overestimation of the magnitude and significance of species-environment associations as well as the artificial inflation of predictive accuracy of the model, which can result in poor predictions in novel environments. Nevertheless, the extent that spatial autocorrelation affects the prediction of the distribution of introduced species is not well understood. To address the influence of spatial autocorrelation on SDM predictions, we compare the results from a spatial model, which controls for spatial autocorrelation, and a traditional non-spatial model developed using native and introduced distribution data from 20 species of North American freshwater fishes. Modeling is carried out using a Bayesian approach that more accurately quantifies the degree of uncertainty in model parameters and predictions compared to traditional frequentist methods.  

Results/Conclusions

By modeling the distribution of native species with established introduced populations, we are able to directly assess the effects of spatial autocorrelation on model predictability. Three general patterns emerged when comparing the results of spatial and non-spatial models: 1) magnitudes of parameter estimates for non-spatial models were greater than spatial models; 2) posterior distributions of predicted probability of occurrence estimates were greater for spatial models compared to non-spatial models; and, 3) non-spatial models consistently over-predicted the introduced distribution of each species. These findings suggest that the significant species-environment associations found in traditional non-spatial models may be, in part, due to spatial patterning in the data. This can result in artificial inflation of the certainty of spatially uniformed model predictions, which can have negative consequences for practical applications of the model. Characterizing the uncertainty associated with predictions from models that incorporate spatial autocorrelation should provide a more realistic reflection of the factors that may regulate the distribution of introduced species.