Exact confidence intervals and hypothesis tests for parameters of discrete distributions (Q502885)

The paper under review discusses confidence intervals and hypothesis tests for univariate parameters of discrete distributions. The authors argue that strictly two-sided confidence intervals and hypothesis tests suffer from problems that are universal when the data is generated by a class of discrete distributions, which includes the binomial, Poisson, and negative binomial distributions. Several known methods for constructing short intervals lack strict nestedness (i.e. accepting a lower confidence level may not lead to a shorter confidence interval). Intervals that are not strictly nested correspond to a class of tests that attach widely different evidence to almost indistinguishable hypotheses. It is shown that among strictly nested intervals, fiducial intervals, including the Clopper-Pearson interval for a binomial proportion [\textit{C. J. Clopper} and \textit{E. S. Pearson}, Biometrika 26, 404--413 (1934; JFM 60.1175.02)] and the Garwood interval for a Poisson mean [\textit{F. Garwood}, Biometrika 28, 437--442 (1936; Zbl 0015.26206)], are optimal.