No label defined (Q4946930)

Fiducial inference is an ``all purpose'' method of making inference, and hence it is natural to discuss its optimality not w.r.t. a particular loss function but w.r.t. the class of all reasonable loss functions. The related concept of strong unbiasedness is so restrictive that it is not to be expected that procedures satisfying this restriction are admissible w.r.t. a particular loss function. An interesting feature of strong unbiasedness, however, is that it recovers classical results like, e.g., Student's inference. Fisher's fiducial argument is a tool for deriving inferences in the form of a probability distribution on the parameter space, not based on Bayes' theorem. Lindley established that in exceptional situations fiducial inference coincides with posterior distributions; in other situations fiducial inference has always been critized. In his comment on \textit{G.N. Wilkinson} [J. R. Stat. Soc. Ser B 39, 119-171 (1977; Zbl 0373.62002)] James wrote: ``Fisher's postulate of ignorance for fiducial theory should be replaced by a positive requirement that the statistical analysis is appropriate to and meets the issues at stake.'' In this paper the properties of fiducial inference are discussed within the limited context of statistical models. Such a positive requirement is obtained from a decision theoretic approach: in some interesting situations procedures provided by the fiducial argument are optimal in a class of risk-unbiased procedures. For the entire collection see [Zbl 0928.00040].