Abstract:

A longstanding conundrum in environmental data with multiply leftcensored measurements (i.e., nondetects) is how to best fit parametric data models. The magnitude and pattern of censoring jointly impact the information available for testing specific models or for identifying outliers. Yet such testing is important for many applications, e.g., estimating limits on contaminants in soil or groundwater. This paper presents a novel strategy for computing percentage points under nearly arbitrary leftcensoring for two common tests: the probability plot correlation coefficient goodnessoffit test and Rosner's block outlier screening test. The new strategy: (1) eliminates the need to 'fudge' percentage points computed from complete samples, (2) allows a unique set of percentage points to be computed for each dataset, depending on the magnitude and pattern of censoring, and (3) partially alleviates the complication of leftcensoring in the 'chickenandegg' problem of needing a distributional model to identify outliers, but also needing to remove outliers prior to fitting a data model.
