Reproducibility and imputation of air toxics data.

Ambient air quality datasets include missing data, values below method detection limits and outliers, and the precision and accuracy of the measurements themselves are often unknown. At the same time, many analyses require continuous data sequences and assume that measurements are error-free. While a variety of data imputation and cleaning techniques are available, the evaluation of such techniques remains limited. This study evaluates the performance of these techniques for ambient air toxics measurements, a particularly challenging application, and includes the analysis of intra- and inter-laboratory precision. The analysis uses an unusually complete-dataset, consisting of daily measurements of over 70 species of carbonyls and volatile organic compounds (VOCs) collected over a one year period in Dearborn, Michigan, including 122 pairs of replicates. Analysis was restricted to compounds found above detection limits in > or =20% of the samples. Outliers were detected using the Gumbell extreme value distribution. Error models for inter- and intra-laboratory reproducibility were derived from replicate samples. Imputation variables were selected using a generalized additive model, and the performance of two techniques, multiple imputation and optimal linear estimation, was evaluated for three missingness patterns. Many species were rarely detected or had very poor reproducibility. Error models developed for seven carbonyls showed median intra- and inter-laboratory errors of 22% and 25%, respectively. Better reproducibility was seen for the 16 VOCs meeting detection and reproducibility criteria. Imputation performance depended on the compound and missingness pattern. Data missing at random could be adequately imputed, but imputations for row-wise deletions, the most common type of missingness pattern encountered, were not informative. The analysis shows that air toxics data require significant efforts to identify and mitigate errors, outliers and missing observations, and that these steps are essential and should be performed prior to using these data in receptor, exposure, health and other applications.