Detection of Outliers in Pearson Type III Data

The Bulletin 17B guidelines for flood frequency analysis provide a method for detecting outliers that is constrained in three ways: (1) It assumes zero skew; (2) it does not address multiple outliers; and (3) it only provides for a 10% level of significance. Because these guidelines also specify a log-Pearson Type III distribution and skews different from zero are common, critical deviates for detecting outliers in samples with skews from −1 to 1 are needed. Using Monte Carlo simulation with Pearson Type III distributions, critical deviates for detecting outliers were computed. The analyses provide functions that describe the relation between the critical deviate and sample sizes from 10 to 150. Critical deviates were developed for detecting up to three outliers from Pearson Type III samples with skews from −1 to 1 and for three levels of significance (10%, 5%, and 1%). The critical deviates are expected to be within 1% of the true values. Flood records from 50 USGS stream gauges were analyzed to assess the effect of using these skew-dependent critical deviates. Comparison of the skew-dependent test to the Bulletin 17B zero-skew test revealed that the zero-skew criterion led to incorrect decisions in 30% of the cases. The importance of the level of significance was also investigated using the 50 flood records. At the 5% level, only one low and three high outliers were detected, whereas at the 10% level, six low and five high outliers were detected.