Testing For Outliers In Radionuclide Data

摘要

The problem of monitoring atmospheric radionuclides over time is investigated. Such monitoring is desirable for both natural and anthropogenic radionuclides. The statistical problem is one of testing for a time series outlier, and the problem is complicated by the fact that often several observations may be missing. In fact it may be the case that several missing observations may occur immediately prior to a data value that is to be tested as an outlier. Evans (1996) proposes an exponentially weighted moving average (EWMA) approach for detecting these outliers. The EWMA approach is one that is quite popular in practice, but it is restricted to some extent by the fact that it is based on the assumption that the autoregressive integrated moving average model, ARIMA(0,1,1), is a good fit to the data. Evans presents simulation results based on simulated radionuclide data obtained from a model that he fit to Kuwait Be7 data consisting of a sinusoidal component with long period plus an autoregressive component. One problem with Evans' approach is that false alarm rates tend to be high when the data value to be tested as an outlier is preceded by a string of missing observations. In this paper we describe several alternative approaches for outlier detection, and we compare these with the Evans method using a simulation study. In this study, outlier detection capabilities are compared in the case in which no data are missing immediately prior to the data value to be tested as an outlier as well as in the more difficult case in which several data values are missing immediately prior to this value. Our results indicate that an autoregressive-based procedure suggested here has much better control over the false alarm rates than does the Evans procedure, and it has detection capability that is comparable to and sometimes better than that obtained by the Evans approach.