ABSTRACT
ABSTRACT Saturated designs are very useful for screening factors and in experiments where the observations are very difficult or expensive to obtain. Due to lack of degrees of freedom available to make inferences about the parameters in the model, the frequentist approach is extremely limited, so the Bayesian approach is a good way to analyse data from saturated designs. However, because of the reliance on prior information, Bayesian methods must be used carefully. Conjugate priors are shown to be somewhat inflexible, whereas priors using finite mixtures of densities yield more natural posterior densities. Also shown is that non-conjugate priors, with independence between the effect parameters and the variance, can be useful.
