Imputation of missing values in a precipitation–runoff process database

Abstract

Hydrologists are often faced with the problem of missing values in a precipitation–runoff process database to construct runoff prediction models. They tend to use simple and naive methods to deal with the problem of missing data. Thus far, the common practice has been to discard observations with missing values. In this paper, we present some statistically principled methods for gap filling and discuss the pros and cons of these methods. We employ and discuss imputations of missing values by means of self-organizing map (SOM), multilayer perceptron (MLP), multivariate nearest-neighbor (MNN), regularized expectation–maximization algorithm (REGEM) and multiple imputation (MI) in the context of a precipitation–runoff process database in northern Iran in order to construct a serially complete database for analyses such as runoff prediction. In our case, the SOM and MNN tend to give similar and robust results. REGEM and MI build on the assumption of multivariate normal data, which we don't seem to have in one of our cases. MLP tends to produce inferior results because it fragments the data into 68 different models. Therefore, we conclude that it makes most sense to use either the computationally simple MNN method or the more demanding SOM.