Problems with outlier test methods in flood frequency analysis

Abstract

In flood frequency analysis, the term “outlier” is commonly used to denote large floods in the systematic record or historical floods which lie far above the majority of the floods in the sample. The mere existence of these outliers complicates the frequency analysis procedure. In order to avoid any subjectivity in the detection and treatment of outliers, the U.S. Water Resources Council (WRC) recommended a method based on the principles of hypotheses testing. In spite of the fact that it has been extensively applied in the United States, there are some theoretical and practical aspects which require further consideration. A study of principles reveals that outlier tests in a statistical context postulate an assumption that outliers have a unique distribution which is different from that of the remaining sample observations. Thus, the theory underlying the outlier test is in conflict with the phenomenon of outliers in flood data because it is generally accepted that both historical floods and extraordinary floods in systematic records all come from a common unknown population including all floods. Consequently it would not be reasonable to introduce outlier tests into a flood frequency analysis. In addition, the so-called “masking effect” encountered in the practical use of the outlier test method in U.S. Water Resources Council (1981) is analytically discussed. Observed flood records at several stations are used to illustrate that the test in WRC Bulletin 17B does not guarantee the detection of outliers if more than one is present in the sample due to this “masking effect”.