LINEAR MODEL DISCRIMINATION THEORY APPLIED TO THE CHOICE OF STRUCTURE AND FORM OF HYDROLOGIC REGRESSION MODELS.

The use of regression models whose coefficients are not fixed but vary randomly was investigated in an attempt to represent uncertainties which are not only additive but also of the multiplicative form. Bayesian theory allows to explicitly account for both parameter and model uncertainties and it was used in this work to derive a procedure to discriminate alternative hydrologic regression models. In particular the procedure was used to discriminate alternative exogenous variables, to compare different structural forms of regression models and different assumptions on the covariance matrix of the distrubances. Artificial and real world examples in water resources were designed to test the proposed procedure under different assumptions on the prior probability of the models, length of the sample, model subset and covariance matrix of the disturbances.