Abstract. Statistical simulation in hydrology is discussed from a Bayesian perspective. The inherent difficulties in both parametric simulation, based on a parent distribution, and classical nonparametric simulation, based on the bootstrap, are discussed. As an alternative to these procedures, a nonparametric Bayesian simulation methodology, P61ya resampling, is introduced. It consists of simulating from a nonparametric predictive distribution obtained from the analysis of a reference sample, and it is asymptotically equivalent to the bootstrap. The method is generalized to take into account a prior hypothesis on the parametric distribution of a variable. A hybrid simulation model is then obtained that includes parametric and nonparametric simulation as particular cases. An extensive application is presented in a related paper [Fortin et al., 1997], where P61ya resampling is used to compare statistical models for flood frequency analysis. In this paper an example is used to demonstrate how P61ya resampling can help assess the influence of a distribution hypothesis on simulation results.
IntroductionIn statistical hydrology, parametric and nonparametric simulation procedures have been extensively used to compare statistical models of hydrological variables. The purpose of this paper is to present an alternative, based on Bayesian analysis, to classical parametric and nonparametric simulation techniques for independent and identically distributed (i.i.d.) variables. The proposed method, which we call P61ya resampling, is similar to the bootstrap [Elton, 1979], which consists in drawing observations with replacement from a reference sample. This simulation scheme, introduced by Lo [1988], is discussed in detail for binomial variables, then for multinomial variables, and finally in a completely nonparametric setting. The methodology is then generalized to take into account prior information independent from the reference sample. The present paper is mainly theoretical; an illustrative example which compares the GEV and Gumbel distributions for at-site flood frequency analysis is included. An extensive application to at-site flood frequency analysis is presented in a related paper [Fortin et al., 1997].
Simulation in Statistical HydrologyA common problem in statistical hydrology is to approximate the unknown statistical distribution F of a (hydrological) random variable X, given prior information and observed data. The usual approach consists in selecting a parametric distribution having probability density function (p.d. In both cases, statistical distributions and methods for estimating the parameters may be compared in terms of their ability to approximate the distribution of the simulated samples. The main problem with parametric simulation is that the parametric distribution of most hydrological variables is unknown and that the choice of the parent distribution usually is arbitrary. Although the robustness of the conclusions to the hypothesis of a given parent distribution may be studied Kuczera, 1982;Haktanir,...