Latent Dirichlet Allocation regression model for biodiversity decomposition

Background/Question/Methods Clustering approaches have been widely used to decompose biodiversity data into communities that have different responses to environmental alteration, but hard clustering approaches fail to represent gradual changes in communities along environmental gradients and poorly represent uncertainty when predicting community distributions at large spatial extents. Latent Dirichlet Allocation model is a kind of mixed membership models and can represent gradual changes in communities. Despite of the advantage of this model, so far no work has been done to incorporate covariate information in model structure, largely limiting the ability of this model in explaining ecological processes and making ecological predictions. In this study, we develop a Bayesian mixed membership model to explain and predict community distributions using environmental covariates. We use simulations to examine the inference of this model, and a case study to illustrate its application.

Results/Conclusions Simulation studies show the capability of our models to select the optimal number of communities and estimate regression parameters governing community-environment relationships. We demonstrate the capabilities of our models with a case study of vascular plants in Alberta in relation to environmental conditions. Our results revealed important impacts of climate, forest type, wetland habitat, and human disturbance on plant community distributions. Furthermore, cross-validation shows that incorporating the information of covariates largely improved the predictive performance of the models. In addition to predicting community distributions, our models are also capable to identify transient areas dominated by multiple communities and predict the distributions of each single species. Because of the important characteristics of our models, we believe that our work is important for both ecological research and conservation practice.