Kernel local partition processes for functional data

David B. Dunson

Department of Statistical Science, Duke University

August, 2008

Functional data analysis commonly relies on the incorporation of basis functions having subject-specific coefficients, with the choice of basis and random effects distribution important. To allow the random effects distribution to be unknown, while inducing subject-specific basis selection and local borrowing of information across subjects, this article proposes a kernel local partition process (KLPP) prior. The KLPP selects the elements in a subject's random effects vector locally from a collection of unique coefficient vectors, leading to a flexible local generalization of the Dirichlet process and to a sparse representation of complex functional data. Basic theoretical properties are considered, an MCMC algorithm is developed for posterior computation and the methods are applied to hormone data.

Keywords: Basis functions; Dirichlet process; Longitudinal data; Nonparametric Bayes; Random effects; Sparsity.


The manuscript is available in PDF formats.