Sharp hypotheses and bispatial inference

Abstract

A fundamental class of inferential problems are those characterised by there having been a substantial degree of pre-data (or prior) belief that the value of a model parameter was equal or lay close to a specified value, which may, for example, be the value that indicates the absence of an effect. Standard ways of tackling problems of this type, including the Bayesian method, are often highly inadequate in practice. To address this issue, an inferential framework called bispatial inference is put forward, which can be viewed as both a generalisation and radical reinterpretation of existing approaches to inference that are based on P values. It is shown that to obtain an appropriate post-data density function for a given parameter, it is often convenient to combine a special type of bispatial inference, which is constructed around one-sided P values, with a previously outlined form of fiducial inference. Finally, by using what are called post-data opinion curves, this bispatial-fiducial theory is naturally extended to deal with the general scenario in which any number of parameters may be unknown. The application of the theory is illustrated in various examples, which are especially relevant to the analysis of clinical trial data.