Multivariate regression trees (MRT) are a new statistical technique that can be used to explore, describe, and predict relationships between multispecies data and environmental characteristics. MRT forms clusters of sites by repeated splitting of the data, with each split defined by a simple rule based on environmental values. The splits are chosen to minimize the dissimilarity of sites within clusters. The measure of species dissimilarity can be selected by the user, and hence MRT can be used to relate any aspect of species composition to environmental data. The clusters and their dependence on the environmental data are represented graphically by a tree. Each cluster also represents a species assemblage, and its environmental values define its associated habitat. MRT can be used to analyze complex ecological data that may include imbalance, missing values, nonlinear relationships between variables, and high-order interactions. They can also predict species composition at sites for which only environmental data are available. MRT is compared with redundancy analysis and canonical correspondence analysis using simulated data and a field data set.
