Predicting spatial patterns of plant species richness: a comparison of direct macroecological and species stacking modelling approaches

Aim This study compares the direct, macroecological approach (MEM) for modelling species richness (SR) with the more recent approach of stacking predictions from individual species distributions (S-SDM). We implemented both approaches on the same dataset and discuss their respective theoretical assumptions, strengths and drawbacks. We also tested how both approaches performed in reproducing observed patterns of SR along an elevational gradient. Location Two study areas in the Alps of Switzerland. Methods We implemented MEM by relating the species counts to environmental predictors with statistical models, assuming a Poisson distribution. S-SDM was implemented by modelling each species distribution individually and then stacking the obtained prediction maps in three different ways – summing binary predictions, summing random draws of binomial trials and summing predicted probabilities – to obtain a final species count. Results The direct MEM approach yields nearly unbiased predictions centred around the observed mean values, but with a lower correlation between predictions and observations, than that achieved by the S-SDM approaches. This method also cannot provide any information on species identity and, thus, community composition. It does, however, accurately reproduce the hump-shaped pattern of SR observed along the elevational gradient. The S-SDM approach summing binary maps can predict individual species and thus communities, but tends to overpredict SR. The two other S-SDM approaches – the summed binomial trials based on predicted probabilities and summed predicted probabilities – do not overpredict richness, but they predict many competing end points of assembly or they lose the individual species predictions, respectively. Furthermore, all S-SDM approaches fail to appropriately reproduce the observed hump-shaped patterns of SR along the elevational gradient. Main conclusions Macroecological approach and S-SDM have complementary strengths. We suggest that both could be used in combination to obtain better SR predictions by following the suggestion of constraining S-SDM by MEM predictions.