Detecting Patterns and Interactions in Ecological Time-series Data Using Model-based Hierarchical Clustering

Edward Baskerville, University of Michigan

Photo of Edward Baskerville

Quantitative models of ecological dynamics are both vital aids to understanding ecological communities and fundamental tools for making predictions about those communities. These dynamics emerge from species' responses to their abiotic environment as well as their interactions with other species. In systems where long time series are available and species interactions are known a priori, simple autoregressive models have proven successful. Where time series are shorter, the number of species is very large and interactions are unknown, the parameter and model space of these models becomes prohibitively large. In this work, simple autoregressive models are combined with a Bayesian hierarchical clustering technique in order to vastly reduce the effective size of the parameter space. Interactions are simultaneously captured using a low-dimensional model parameterization. The resulting fitted model provides a biologically interpretable description of the ecological community in terms of a hierarchy of functional groups, their responses to each other, and their responses to environmental covariates. Additionally, this method naturally extends to other kinds of data and model forms.

Abstract Author(s): Edward B. BaskervilleEoin L. BrodieAdam P. Arkin