Correcting bias in the fossil record to improve pooled niche estimation and species distribution model transferability
Paleorecords and species distribution models (SDMs) are increasingly combined to improve our understanding and forecasts of biotic response to climate change. To account for changes in species-climate relationships and to better reconstruct species’ fundamental niches, it has been suggested that fossil data and associated environmental variables should be pooled across multiple time periods (pooled niche; PN). Pooling data is especially feasible for the late Quaternary given the high spatiotemporal density of fossils that document pronounced changes in species assemblages (and niches) across the large and rapid climate changes during the last deglaciation. However, because the quantity of fossil data decreases back in time and there are long periods where temperature variations are relatively small (i.e., the Holocene), naïve pooling may bias inference of the PN towards the Holocene. Using observed changes in distribution of 19 plant taxa (recorded in fossil pollen) in eastern North America and CCSM3 paleoclimate simulations, we studied the effect of (1) the quantity of fossil pollen records and (2) greater representation of Holocene climates on PN-SDMs. Furthermore, we tested their ability to forecast species distributions by projecting the PN models to multiple time periods and comparing them with models fit at single time periods.
We found that both types of bias (i.e. sample size and stable temperatures in the Holocene) strongly affect PN estimates towards Holocene conditions. Niche overlap analyses between single periods and partitioning of the contribution of each time period to the PN show that data from the Last Glacial Maximum and the early-Holocene transition provide the most unique niche information, providing valuable information of suitable (or at least conditions that allow persistence) climatic conditions. However, these biases can be corrected with subsampling and niche overlap analyses, which essentially remove redundant information. Overall, fitting SDMs with PN information improved model transferability across time. Unbiased PN models (i.e., models fit pooling data but controlling for the sample size and stable temperatures in the Holocene) provided the best performance across all time periods. Finally, we found that PN models tend to improve prediction of presence but not absence. This result supports the idea that PN models are estimating the potential distribution and that some taxon absences may reflect the effect of non-climatic drivers not directly related to the physiological limits of the taxa such as biotic interactions, dispersal limitations or population dynamics.