@BLOCK CENTER
 DATA FROM CH. 16 OF AFIFI & CLARK
 EXAMPLE WITH N=5
 
@END
    MIXPCMA CLUSPAC - MIXTURE-MODEL CLUSTERING OF CASES
    COMMON COVARIANCE MATRIX
                                   MODEL-SELECTION CRITERIA
     NUMBER     MINUS 2    NUMBER  (PPH=POST.PROB.OF MODEL):
       OF        TIMES      OF      AKAIKE         SCHWARZ
    CLUSTERS,    LOG OF    PARAM- ------------   ------------
        K      LIKELIHOOD  ETERS   VALUE   PPH    VALUE   PPH
    ---------  ----------  -----  -------  ---   -------  ---
        1          43.6       5      53.6 0.99     51.7  0.09
        2          42.3       8      58.3 0.01     55.2  0.02
        3          42.0      11      64.0 0.00     59.7  0.00
        4          24.1      14      52.1 0.00     46.7  0.89
        5          34.4      17      68.4 0.00     61.8  0.00
 
    WITH K = 2, THE CLUSTERING WAS {1,2,3},{4,5} RATHER THAN THE
    OBVIOUSLY "CORRECT" {1,2},{3,4,5}.
    .................................................................
 
    MIXPDTA CLUSPAC - MIXTURE-MODEL CLUSTERING OF CASES
    VARYING COVARIANCE MATRICES
                                   MODEL-SELECTION CRITERIA
     NUMBER     MINUS 2    NUMBER  (PPH=POST.PROB.OF MODEL):
       OF        TIMES      OF      AKAIKE         SCHWARZ
    CLUSTERS,    LOG OF    PARAM- ------------   ------------
        K      LIKELIHOOD  ETERS   VALUE   PPH    VALUE   PPH
    ---------  ----------  -----  -------  ---   -------  ---
        1          43.6       5      53.6 0.13     51.7  0.00
    ==> 2          23.5      11      45.5 0.87  ==>41.2  0.97
        3          26.5      17      60.5 0.00     53.8  0.03
        4          32.7      23      78.7 0.00     69.7  0.00
        5          43.6      29     101.6 0.00     90.3  0.00
 
 
    WITH K = 2, THE CLUSTERING WAS THE OBVIOUSLY "CORRECT"
    {1,2},{3,4,5}.
    .................................................................
 
 
     Performance of the default option of BMDPKM
(Dixon 1990) on these data:
 
(Afifi and Clark 1990, p. 445)
The program starts with all cases in a single cluster.  The
program then searches for the variables with the highest
variance, in this case the first variable.  The original cluster
is then split into two clusters, using the midrange of the first
variable as the dividing point.  After this every case already belongs
to the
cluster whose centroid is closest to it.  Thus the algorithm
stops, with the two clusters selected being clusters {1,2,3} and
{4,5}.
 
Dixon, W.J., ed. (1990).  BMDP Statistical Software 1990.  Vols. 1 and 2.
Univ. of Calif. Press, Berkeley.