Fiducial inference in the pivotal family of distributions (Q862687)

The aim of this paper is to give a general method of deriving fiducial distributions, to show the frequentist properties of fiducial distributions, and to exploit the flexibility of dealing with various statistical inferences. The proposed method is based on three concepts: pivotal families of distributions, generalized pivotal models, and fiducial models. A family of distributions is called pivotal if a generalized pivotal model can be found. Analytical results show that a large number of parametric families of distributions are pivotal. R.A. Fraser (1930) was the first to introduce the fiducial method and to give a fiducial inference in the transformation families of distributions, where the fiducial model plays an essential role. A fiducial model is a function of a random variable with a known distribution, called the pivotal random element, with the observation of a statistic being given. The proposed method for the device of pivotal families of distributions encompasses some other methods of deriving fiducial distributions. Especially the first fiducial distribution given by R.A. Fisher can be deduced within the proposed framework. For the monotone likelihood ratio family of distributions, which is a pivotal family, the fiducial distributions are proved to yield a frequentist property in the Neyman-Pearson view, i.e., the confidence bounds of parameters determined from the Fisher's fiducial distributions are the uniformly most accurate bounds. Moreover, some statistical inferences such as the confidence region of parameters, testing hypotheses and variable predictions can be fulfilled by the fiducial models combined with the generalized pivotal models. Four applications of the fiducial distributions point out the advantages of the fiducial inference. Several examples are exposed illustrating the proposed method, including examples in which the fiducial distributions cannot be obtained by the existing methods.