Robust flood frequency models

The concept of a robust model is briefly explored. In the context of flood frequency analysis, two necessary properties of a robust model are advanced, namely, resistance and efficiency. Strategies for seeking more robust models are discussed. Because of its versatility, the five‐parameter Wakeby distribution can credibly be considered a parent flood distribution. Four regionalized Wakeby parents are employed in simulation studies to search for robust models. These parents were shown by Houghton to be representative of U.S. flood experience in the sense that certain raw flood data characteristics could be reproduced. A limited range of sampling experiments were undertaken. The results suggest that of the site‐specific estimators considered, the two‐parameter log normal maximum likelihood estimator is most resistant, with Gumbel estimators employing either maximum likelihood or probability‐weighted moments displaying comparable resistance. Several estimators which utilize regional flood information were compared. Included were empirical Bayes estimators which are structurally similar to James‐Stein rules and regionalized estimators based on the flood index method. These estimators exhibited substantial improvements in aggregate risk performance over their site‐specific counterparts, particularly for short record lengths. Regionalized estimators appear to be preferable for short record lengths, while estimators which combine both site and regional flood information are preferable for longer record lengths. When such estimation procedures are considered, other distributional models such as log Pearson type III and Wakeby become practical alternatives to the two‐parameter log normal model.