A Bayesian Approach to the Design and Analysis of Experiments for Regression Models,

A Bayesian approach to the design and analysis of experiments for linear regression models is presented, where the objectives of the experiment are satisfied by a joint design criterion reflecting concern for both model discrimination and parameter estimation. Under the assumption of unknown variance, a probability mixture representing the state of the system is formulated and the procedure sequentially selects design points which maximize the posterior marginal variance of the response surface. Several stopping rules for termination of the experiment are proposed and a number of simulations illustrating the use of this procedure are included. Some advantages of this procedure are that it is easily implemented as an on-line controller and allows the experimenter maximum flexibility in allocating resources and deciding when to terminate experimentation. Author