Vegetation changes during the late Quaternary: Predicting ability of community level models across time

Background/Question/Methods

Predicting the response of biotic systems to environmental change remains one of the greatest challenges in ecology. During the last decade, studies have emphasized the use of species distribution models (SDMs) to predict climate-driven shifts in species distributions and extinction risk. SDMs usually are fit using only abiotic factors, even though other factors can modify the strength and/or the direction of abiotic drivers. Among these factors, biotic interactions are known to play a key role in determining species’ responses to climate change but the implementation of such interactions in SDMs remains limited. Alternative approaches to SDMs are multivariate tools that simultaneously model community structure and composition. These community level models (CLMs) can provide more reliable predictions for species distributions and community structure by accounting for patterns of co-occurrence (and ostensibly biotic interactions) between species in the model. However, this capacity remains largely unexplored. Using observed changes in plant associations (as recorded in fossil pollen records) in eastern North America and independent paleoclimate simulations, we tested the ability of five CLMs to forecast species distributions, species associations, and macroecological patterns across time. Specifically, we fit CLMs with current presence-absence data and hindcasted at 500-year intervals until 21 ka BP.

Results/Conclusions

Among the five CLMs, vector generalized additive model (VGAM) and vector Generalized Linear Model (VGLM) best predicted pollen taxon distribution, community composition and taxon richness (mean AUC: 0.85; mean Jaccard index between observed and predicted communities: 0.42; mean correlation between predicted and observed species richness: 0.75). In contrast, neural network (NNET) and classification and regression trees (CARTs) performed the worst (AUC: 0.75; Jaccard index: 0.55; species richness correlation: 0.5). However, multivariate adaptive regression spline (MARS) and NNET offered the best estimates of beta diversity (Sorensen index). All CLMs predictions were independent of species prevalence and their predictive ability decreased backwards through time. Projections remained reliable (AUC > 0.75) until the mid-Holocene (7 ka BP), decreased from 7 to 11 ka BP, and then remained very unreliable (AUC ~ 0.6) through the whole Pleistocene (from 11 ka to 21 ka BP). Our results show that predictive performance of CLMs differed between algorithms and consistently declined as climate and community dissimilarity with present increased. Ongoing research is comparing these results to SDMs to determine the extent to which CLM may complement, or represent an alternative to, species-level modeling.