Outlier Detection in Multivariate Hydrologic Data
Abstract
The existence of extreme events in data sets can distort the accuracy of computed statistics. Some favor censoring outliers while others oppose censoring measured values. Detection methods are available for use in the univariate case, but methods for identifying outliers in multivariate data are limited. Past research has provided detection methods, but they are limited by the number of outliers, the number of predictor variables, or limited levels of significance. A method for detecting outliers in multivariate data that is valid for as many as five outliers is presented. Critical values for levels of significance for sample sizes from 10 to 100 are presented. The method is tested on actual hydrologic data. The method is compared with Rosner’s univariate outlier test. The differences between the two methods, the effectiveness of the two methods, and their limitations are examined.