ASSESSMENT OF PRIOR DISTRIBUTIONS FOR REGRESSION-MODELS - AN EXPERIMENTAL-STUDY

Kadane et al. (1980) describe a method of eliciting a subjective conjugate prior distribution for a linear regression model with the usual normal error structure. Here we describe an experiment in which assessors used a modified form of the method to quantify their opinions about four linear models. The accuracy of assessed prior distributions is measured by a scoring rule and different ways of implementing the method are considered and the effect on scores examined. The empirical results provide guidance for using the method in practice. In addition, the information contained in assessed prior distributions is quantified in terms of equivalent sample sizes, using methods developed here. The question of whether assessed distributions are useful rather than misleading is also addressed.