Bayesian design for Copula models
Student: Jagath Senarathne
When: Wednesday, 6th March 2017 2:00 PM-3:00 PM
Where: GP-Z Block, Level 3, Room 305
- A/Prof James McGree (Principal)
- Dr Christopher Drovandi
- James McGree (Chair)
- Tony Pettitt
- Peter Baxter
- Christopher Drovandi
There are many situations where multivariate responses are measured during an experiment, such as drug toxicity and efficacy studies, and experiments for yield and disease resistance of plants. However, design of experiments for multivariate responses has not received much attention particularly when compared to univariate responses largely due to appropriate multivariate distributions not being available.
Copula models provide flexible alternative structures to derive the joint distribution of multivariate responses. However, they are rarely considered in the experimental design context, particularly in a Bayesian framework where model and parameter uncertainty are considered. Thus, new approaches are required to consider optimal experimentation under this framework.
In this work, we aim to develop new Bayesian computational algorithms for deriving optimal designs for experiments with multivariate responses described by Copula models. We explore a variety of such models which explain dependence structures in experiments where bivariate discrete and mixed responses are observed. The sequential Monte Carlo algorithm is adopted to reduce the computational effort required in deriving sequential designs efficiently. Designs are derived based on a dual purpose (model discrimination and parameter estimation) utility function, and compared against designs for single objectives.