Ecology is centered on the relationship between organisms and their environment, which depends on the location of the organism. Distribution modeling uses the location of organisms to come to conclusions about their ecology and potential distribution. Often this approach is based on presence-only or presence/absence data. However, we know that species can occur in locations where they cannot survive in the long-term (e.g. sink populations) and that the occurrence of a single individual is not ecologically equivalent to the occurrence of a large population. I investigated the consequences of relying on presence-only information to model selection, model performance, and ecological insights gained from the models in contrast to using full abundance information in a simulation model. Specifically, I simulate a species that exhibits complications such as spatial autocorrelation, latent variables, limiting effects, and dependence structures among known and latent variables. Then I test the ability of different modeling techniques to recover the processes that led to the species’ distribution contrasting their relative success based on having species abundance vs. having species presence only.
Results/Conclusions
Presence-only information led to a predictable decrease in performance of all models. This decrease in performance was exacerbated when the distributions depended on complicated combinations of variables, limiting effects, latent variables, and intercorrelations among variables. In particular, automated variable selection suffered when ecological situations were complicated and when the dependent variable was a simplistic presence-only. In addition, the gap between goodness-of-fit and predictive performance of models widened, meaning the models were more overfit, as the system became more complicated. The lesson for ecology is that complicated models generally cannot compensate for ecologically insufficient information. Simply knowing the presence of an organism is insufficient to gain rigorous insights into the species’ niche and distribution. When studies concluded that presence-only in conjunction with advanced models are sufficient to achieve good fits or predictive performance, the evaluation methods were typically flawed.