Foundations of Statistical Inference II.

摘要

Some principal technical developments of Part I of this paper are derived here in more elementary fashion, under tho restriction to statistical experiments with discrete sample spaces, but under the more general condition that any finite number of (simple) statistical hypotheses may be represented. For any such experiment, it is shown that for typical purposes of Informative statistical Inference, just the likelihood function on an observed outcome can and should be reported and Interpreted to provide inferences of general Interest concerning the statistical hypotheses (or unknown parameter values); and that for such purposes, the structure of the experiment from which an outcome was obtainedis Irrelevant, apart from determination of the likelihood function. Specific techniques for interpretation of likelihood functions are developed, particularly 'intrinsic confidencemethods' which constitute an appropriate generalization and refinement of confidence methods and conditional confidence methods. The relations of such methods to traditional methods based on the 'principle of insufficient reason', are discussed, as to form and interpretation. In Sections 9-11, analogous developments are given for experiments involving translation or scale parameters.