Linking relative suitability to probability of occurrence in presence-only species distribution models: Implications for global change projections

Our ecological understanding of how biodiversity will respond to global change is to a large extent based on projections from presence-only species distribution models. Despite the incredible utility of these models, we know that the predictions they generate can be heavily influenced by user decisions about model structure or parameter choices. Here, we test how the function used to convert relative suitability to probably of occurrence in presence-only species distributions models can affect predictions of both the magnitude and location of biodiversity change. We used MaxEnt models to create maps of relative suitability for 354 avian species under both current climate conditions and climate conditions in the year 1981. In a back-casting analysis we tested how well three functions relating relative suitability to probably of occurrence perform in recovering observed changes in range size: (1) a logistic curve with informed, species-specific, prevalence values, (2) the default logistic curve, or (3) a commonly used statistical threshold. We then quantified the implications of these functions for projections of species' future range shifts with climate change. We found that using either the default logistic function or a common threshold function for habitat suitability tends to (1) estimate larger effects of past climate change on species range size than observed in the time-series data, (2) inflate projections of how much future climate change will impact species range size and (3) potentially misidentify the locations of greatest range expansion or contraction. We further provide a mathematical basis for these biases, suggesting their general applicability to other systems. Last, we show that these biases can be avoided by analysing proportional rather than absolute changes in range with climate change, and by abandoning the use of habitat suitability thresholds. Incorporating these practices can facilitate a more predictive use of species distribution models when forecasting the response of biodiversity to global change.