University of Washington - Department of Statistics
We consider nonparametric modeling of within- and between-subject variation. Each subject within a population has multiple observations, distributed according to a subject-specific distribution. Subject-specific distributions are assumed to be "similar" to one another, in some regard. We explore the use of Dirichlet priors and Polya urn schemes to model such a hierarchy, and discuss the deficiencies of such an approach. We suggest an alternative model, based on a smoothness constraint imposed by a mixture of distributions. The work is motivated by a study on the locations of tumors along the mammalian intestine.