Seminars

Many experiments involve binary or count responses that don't follow a normal distribution.  Great advances have been made for modelling the data from these experiments using
Generalized Linear Models (GLM's). Yet we know little about how to design these experiments efficiently and what we know is largely limited to one-factor designs.  For example, classical two-level factorial designs can be very inefficient in binary response
experiments. An important, and problematic, aspect is that good
designs depend on knowledge of the unknown model parameters.  We will present some effective, and practical, algorithms for designing experiments that achieve a high degree of robustness to initial assumptions about the model parameters.  We consider both static experiments, in which all runs are determined at the outset, and sequential experiments, in which the data collected thus far are used in planning subsequent observations. The methods have a Bayesian flavor in that they exploit prior belief about the coefficient values. Our algorithm for static experiments takes advantage of a fast scheme for generating locally optimal designs together with a clustering procedure.  Our algorithm for the sequential case uses a sampling and weighting approach, rather than direct computation, to represent the posterior.  We illustrate the performance of the algorithm on several applications.  For the sequential case, we compare the results from our algorithm with those from the classic ``Bruceton'' method on an actual sensitivity test conducted recently at an industrial plant.