Multivariate trees for mixed outcomes

In this paper, we propose a tree-based method for multivariate outcomes consisting in a mixture of categorical and continuous responses. The split function for tree-growing is derived from a likelihood-based approach for a general location model (GLOM). One situation where the new approach should be appealing is when mixed types of multiple outcomes are used as surrogates for an unobserved latent outcome. Two illustrations of the application of the new method are given with health care data.