|
Abstract:
|
In some toxicology studies, one may expect changes in the shape and location of an outcome distribution with dose. Unfortunately when the data consist of observations on hundreds-thousands of cells or small organisms, it can be difficult to fit models which are sufficiently flexible. To address this problem, we propose a computationally fast method for testing for distribution changes across an ordinal predictor. Our method is based on a Variational Bayes (VB) approach to fitting a Dynamic Mixture of Dirichlet Processes. Our approach begins by approximating the posterior in the lowest predictor group and then, using EM sub-steps, we successively add atoms to the later groups until there is no benefit in terms of the approximated marginal likelihood. Since VB approaches converge in relatively few steps, our method should be more computationally feasible than MCMC for large data sets.
|
