Using principal component analysis to detect outliers in ambient air monitoring studies

The need to determine outliers in analytical data sets is important to ensure data quality. More sophisticated techniques are required when the checking of individual results is not possible, for instance with very large data sets. This paper outlines a novel method for the detection of possible outliers in multivariate sets of air quality monitoring data, here the metals content of ambient particulate matter. Principal component analysis has been used to take advantage of the expected correlation between metals concentrations at individual monitoring sites to produce a summary statistic based on the deviation of each observation from the expected pattern, which can then be interrogated using one-dimensional robust statistical techniques to identify possible outliers. The sensitivity of this statistic to the number of principal components included in the summary statistic has been examined, and the method has been demonstrated on exemplar data from the UK Heavy Metals Monitoring Network where it has produced very accurate predictions of outlying data.