40 Outlier Identification in Censored Environmental Time Series

Oregon Convention Center 

When estimating trends in contaminated media, it is common to jointly observe apparent outliers and non-detects (i.e., left-censored observations). Identifying outliers usually requires de-trending the time series prior to screening for outlying residuals. The screening in turn requires a reference distribution from which to judge outlying points.

The combination of censored data, nonlinear trends, and outliers raises challenges: 1) how to estimate the trend prior to treating non-detects, and vice-versa? 2) how to compute 'censored residuals' from the trend? 3) how to build a reference distribution given substantial censoring?

We formerly proposed a Monte Carlo mixture model that samples non-detects from a class of bounded distributions on the interval (0, DL), where DL is a left-censoring limit. We illustrate how this mixture model can accurately identify outliers by constructing a broad trend using the mean estimate of repeated draws from the mixture model, and studentizing the trend residuals to both flag and down-weight outliers via an appropriate kernel applied to the studentized distance from the trend.

The benefits of this strategy are explored.

Outliers

Left-Censored Data

Time Series

Environmental

Monte Carlo

Trends 

Section on Statistics and the Environment