Outlier Detection and Editing Procedures for Continuous Multivariate Data
Published in: Journal of Official Statistics, v. 22, no. 3, 2006, pp. 487-506
Posted on RAND.org on December 31, 2005
In large datasets, outliers may be difficult to find using informal inspection and graphical displays, particularly when there are missing values. We present a semi-automatic method of outlier detection for continuous, multivariate survey data that is designed to identify outlying cases and suggest potential errors on a case-by-case basis, in the presence of missing data. Our method relies on an explicit probability model for the data. The raw data with outliers is described by a contaminated multivariate normal distribution, and an EM algorithm is applied to obtain robust estimates of the means and covariances in the presence of missing values. Mahalanobis distances are computed to identify potential outliers and offending variables. The procedure is implemented in a software product, which detects outliers and suggests edits to remove offending values. We apply the algorithm to preliminary body-measurement data from the Third National Health and Nutrition Examination Survey, Phase I (1988'1991). This method works quite generally for continuous survey data, and is particularly useful when inter-variable correlations are strong.