Model uncertainty in flood frequency analysis and frequency‐based design

Constrained by computational feasibility, attempts to describe random natural phenomena of complex origin analytically can lead to a multiplicity of simplistic potentially representative model forms, as has occurred in the case of flood frequency analysis. Classical statistical methods inadequately confront this model uncertainty. Likelihood and Bayesian methods are presented and shown to permit inference concerning the relative goodness of several potential model candidates with respect to a given set of flood events. The Bayesian inferences are further combined within a decision theoretic structure for examination of the anticipated economic consequences of model uncertainty regarding decisions concerning flood protection levels. Results show that Bayesian methods supply more precise information but require greater effort. Since a model world of simplistic forms may never be defined absolutely, both decision and inference remain subject to the astute judgment of the analyst.