Get access

Modeling the climatic drivers of spatial patterns in vegetation composition since the Last Glacial Maximum


J. L. Blois, Dept of Geography and the Center for Climatic Research, Univ. of Wisconsin – Madison, Madison, WI 53706, USA. E-mail:


Projecting the future composition and function of communities is a major challenge, and there is an urgent need to develop, improve, and test the predictive capacity of ecological models under different climate states. We tested the effect of climate on spatial patterns of plant community composition over the past 21 000 yr, focusing on whether the spatial relationships between environmental distance and compositional dissimilarity are stable over time. We used a network of fossil-pollen sites in eastern North America, combined with paleoclimate simulations from the Last Glacial Maximum (LGM; 21 000 calibrated years before present, 21 kyr BP) to the present. We modeled relationships between climate, geography, and compositional dissimilarity at 1 kyr periods using generalized dissimilarity modeling (GDM) and determined the strongest predictors of compositional dissimilarity. We assessed the performance of models calibrated for one time period (e.g. 14 kyr BP) in predicting patterns in the same period as well as at other times (e.g. 12 kyr BP), and tested whether predictive performance was related to the magnitude of climate change between the calibration and prediction time periods. Finally, we examined whether pooling data from multiple time periods improved predictive performance. Models explained 32 to 51% of compositional dissimilarity between locations within any single time period. The best set of predictors changed across time, with summer temperature and geographic distance the strongest predictors of compositional dissimilarity for most time periods. Models built for one time period explained turnover during nearby time periods relatively well, but performance decayed across time and with increasing climate change. Results were similar regardless of whether models were projected forward or backward through time, and did not improve when data were pooled across time. GDM predicts well the spatial patterns of past compositional dissimilarity and holds promise for modeling the drivers of compositional dissimilarity across space and time. However, the modeled relationships between compositional turnover and environmental distance are non-stationary, so caution is needed when predicting across periods of significant climatic change.