Estimation of rare errors using expert judgement
Hirokuni Tamura · Biometrika · 1988
Cox & Snell (1979) investigate inference about the total population error when most observations are zero. Their approach is parametric and assumes that errors are positive. Empirical evidence indicates, however, that negative errors are not uncommon and that the error distribution is unlikely to be modelled by any standard parametric distribution. We propose a more general nonparametric approach that uses Ferguson's Dirichlet process. Our model allows the investigator to specify a wide range of prior prediction ranging from a standard parametric distribution to observations without any formal structure. Numerical examples are presented for demonstration of this flexibility in modelling. Comparisons with parametric models are discussed.