ON ROHLFS METHOD FOR THE DETECTION OF OUTLIERS IN MULTIVARIATE DATA

Rohlf (1975, Biometrics 31, 93-101) proposed a method of detecting out liers in multivariate data by testing the largest edge of the minimum spanning tree. It is shown here that tests against the gamma distribut ion are extremely liberal. Furthermore, results depend on the correlat ion structure of the data if Euclidean distances are used. While the u se of generalized distances might avoid this difficulty, the construct ion of the robust estimates required to carry out the test with genera lized distances provides in itself information on outliers which leave s Rohlf's procedure superfluous. It is concluded that Rohlf's method d oes not provide a useful formal test. (C) 1995 Academic Press,Inc.