Informal likelihood measures in model assessment: Theoretic development and investigation

Within hydrology performance criteria such as the Nash–Sutcliffe efficiency have been used to condition the parameter space of a model. Their use is motivated by the fact that the stochastic error series between a model output and corresponding observations is the result of the composite effect of multiple error sources which cannot be described, even in form, a priori. This paper formalises the use of such performance criteria within a Bayesian framework, such as Generalised Likelihood Uncertainty Estimation (GLUE), by introducing the concept of Informal Likelihoods. Informal Likelihoods are used to characterise desirable features in the relationship between the model output and corresponding observed data. A number of common performance criteria are considered as Informal Likelihoods. Analytical results and a simulation indicate all of the performance criteria considered as Informal Likelihoods in this paper have one or more properties which may be considered undesirable, but may perform no less well in conditioning model parameters than formal likelihoods for which the assumptions are only mildly incorrect.