[R] plot hclust - canberra dist + median linkage

Martin Maechler maechler at stat.math.ethz.ch
Tue Oct 12 17:38:38 CEST 2004

>>>>> "Dan" == Dan Bolser <dmb at mrc-dunn.cam.ac.uk>
>>>>>     on Mon, 11 Oct 2004 16:21:53 +0100 (BST) writes:

    Dan> Gives strange results.

    Dan> I get 'weird' dendrograms with canberra / binary distance metric and
    Dan> median / centroid cluster methods.

it doesn't depend on the metric: 
Both 'median' and 'centroid' methods are known to *not*
guarantee ``monotone distance measures'', or equivalently to
possibly lead to dendrograms with so called ``inversions''.
We should add this to help page for hclust().

Probably for this reason, agnes() from the cluster package
doesn't have these two methods [explicitly] -- though it now
allows general parameter Lance-William formula methods which can
also lead to inversions.

    Dan> Is this just my data?

evidently not.  Though the problem does not appear for all data

Martin Maechler

More information about the R-help mailing list