SRCOS SESSION    "Recent Advances in Outlier Detection"
	D. Rameriz & D. Scott, Organizers


Title:  "Outlier Detection by a Minimum Distance Criterion"

	David W. Scott
	Dept of Statistics
	Rice University


The detection of outliers is closely linked to the
problem of robust estimation.  Minimum distance
procedures have been shown to be inherently robust
but have not been easy to apply to high-dimensional
data.  We have recently developed a minimum-distance
estimation criterion based upon integrated squared
error.  In this paper, we investigate its utility
as a robust estimator and as a preprocessor for
determining outliers.  We show that this approach
extends the idea of fitting a mixture of normals
to data, and then identifying as outliers those points
in some of the small mixture components.  Application
to regression outlier detection is also considered.