Evaluation of Data with Systematic Errors

Application-oriented evaluated nuclear data libraries such as ENDF and JEFF contain not only recommended values but also uncertainty information in the form of "covariance" or "error files." These can neither be constructed nor utilized properly without a thorough understanding of uncertainties and correlations. It is shown how incomplete information about errors is described by multivariate probability distributions or, more summarily, by covariance matrices, and how correlations are caused by incompletely known common errors. Parameter estimation for the practically most important case of the Gaussian distribution with common errors is developed in close analogy to the more familiar case without. The formalism shows that, contrary to widespread belief, common ("systematic") and uncorrelated ("random" or "statistical") errors are to be added in quadrature. It also shows explicitly that repetition of a measurement reduces mainly the statistical uncertainties but not the systematic ones. While statistical uncertainties are readily estimated from the scatter of repeatedly measured data, systematic uncertainties can only be inferred from prior information about common errors and their propagation. The optimal way to handle error-affected auxiliary quantities ("nuisance parameters") in data fitting and parameter estimation is to adjust them on the same footing as the parameters of interest and to integrate (marginalize) them out of the joint posterior distribution afterward.