Development of an Influence Statistic for Outlier Detection With Time Series Traffic Data.

As part of a SERC funded project investigating the detection and treatment of outlying time series transport data, the practical applicability of the Influence Statistic described by Watson et al(1991) is assessed here. Missing or outlying data occur in a variety of transport time series such as traflic counts or journey times for many reasons including broken machinery and recording errors. In practice such data is patched largely by subjective opinion or using simple aggregate methods. In the analysis of non-transport time series several methods have been recently developed to both detect and treat outliers, including work by Kohn and Ansley (1986), Hau and Tong (1984) and Bruce and Martin (1989). These methods use either an intervention modelling approach (where the outlier is modelled as part of an ARIMA structure) or look at the influence an observation exerts on a particular parameter associated with the model. An alternative is the Influence Statistic proposed by Watson (1987) and Watson et al (1992) which examines the influence of an observation on the sample autocorrelation function. Initial research showed the statistic has practical application in a transport context, and a replacement procedure based on the method was found to be effective in treating maverick data. Here we report the results from a wider application of the statistic using traffic count data fmm. the Department of Transport. Further developments are suggested and investigated for the replacement procedure and a comparison is made between possible variations in the method.