Help System (web edition)

DCLUSTER <data>,<distance_matrix>,L
performs cluster analysis by different methods.
By default it is done by means of `medoids'.
This technique is presented by Kaufman and Rousseeuw in 1987.
See their book `Finding Groups in Data' (Wiley 1990).
It seems to be more robust than the standard k-means method.

The computations in DCLUSTER are based entirely on a ready-computed
<distance_matrix> (obtained in Survo by the DIST operation, for
example) from Survo <data>.
The number of clusters is given by specification GROUPS.
Default is GROUPS=2. The group indices 1,2,... are saved in <data> 
as a variable given by mask `G' and the (optional) `silhouette'
as values of variable with mask `S'.
The greatest possible number N of observations depends on the size of
the central memory of the computer (NxN distance matrix has to be
present). For example, in a 64MB memory N is about 2500.
However, to large data (say, more than 1000 observations) the medoid
method can be applied by taking random samples. This technique is
supplied by the /CLARA sucro.

DCLUSTER <data>,<distance_matrix>,L
with specification METHOD=3 makes cluster analysis
by the single linkage (nearest neighbour) method.
The setup is otherwise similar as described above.
To avoid certain weaknesses of this method like
the chaining effect the distance between two groups
of sizes n1 and n2 is multiplied by
1+weight*[log2(min(n1,n2)+1)-1]  (suggestion of S.M. 1998)
by using the specification WEIGHT=weight. Default is WEIGHT=0
(i.e. standard single linkage) but WEIGHT=1 is recommended.

In METHOD=3 an initial grouping 1,2,... (obtained by methoids)
can be given by (another) grouping variable given by
INIT=<name_of_variable>.
In practice, when the expected number of cluster is, say, 2 or 3
one may start by creating the medoid solution with, say, 10 groups
and then the final solution is found by METHOD=3 with WEIGHT=1
on the basis of this preliminary 10 group solution.
  C = More information on cluster analysis 
  M = More information on multivariate analysis