Application of Variance Ratio Criterion (VRC) by Calinski

and Harabasz (1974)

Calinski and Harabasz (1974) introduced the variance ratio criterion (VRC), which
can be used to determine the “correct” number of clusters in a cluster analysis and
which has proven to work well in many situations (Milligan and Cooper 1985).
For a solution with N objects and K segments, the criterion is given by

VRCk ¼ ðSSB =ðK  1ÞÞ=ðSSW =ðN  KÞÞ;

where SSB is the overall between-segment variation and SSW the overall within-
segment variation with regard to all clustering variables. The criterion should look
familiar, as this is actually the F-value of a one-way ANOVA with K representing
the number of factor levels. Consequently, the VRC can easily be computed using
SPSS, even though this is not readily available in the clustering procedures’ SPSS
outputs
To finally determine the “correct“ number of segments, we compute oK for each
segment solution as follows:

ok ¼ ðVRCkþ1  VRCk Þ  ðVRCk  VRCk1 Þ:

.
In the next step we choose a value for K, which will minimize the value of oK.
Owing to the term VRCK-1, the minimum number of clusters that can be selected is
three. This is a clear disadvantage of the criterion.
We want to illustrate the application of the VRC using the cars.sav dataset. Open
the dataset and go to Analyze ▸ Classify ▸ K-Means Cluster. This displays a new
dialog box (Figure A9.1).

1
2 Application of Variance Ratio Criterion (VRC) by Calinski and Harabasz (1974)

Figure A9.1 k-Means Cluster Analysis Dialog Box.

As with the examples discussed in Chapter 9, move the six clustering variables,
i.e. moment, width, weight, trunk, speed, and acceleration into the Variables box
and specify the case labels (variable: name). Instead of using the cluster centers
from our previous hierarchical cluster analysis, we allow SPSS to randomly select
the initial cluster centers.
As we are interested in comparing different segment solutions, we have to run
the analysis for different numbers of segments. In this case, we want to compute
VRC values for a three-, four-, and five-segment solution. Please note that this
analysis is for illustrative purposes only, as it is not really meaningful to use these
numbers of segments with such low numbers of objects in the dataset. Since
determining a suitable number segments using VRC involves comparing the VRC
values of solutions with one segment less than K and with one segment more than K,
we need to run k-means for a three- through six-segment solution.
We start off by running k-means for a two-segment solution. Enter 2 in the Number
of Clusters box (Figure A9.1). Before starting the analysis, you have to request an
ANOVA table by clicking on Options, which will produce the dialog box shown in
Figure A9.2. The resulting output provides the basis for the computation of VRC.
Application of Variance Ratio Criterion (VRC) by Calinski and Harabasz (1974) 3

Figure A9.2 K-means Options Menu.

This analysis produces a number of outputs such as the initial and final cluster
centers. In addition to the outputs discussed in Chapter 9, SPSS generates an
ANOVA table shown in Table A9.1. The ANOVA table indicates whether there
are significant differences in the clustering variables across the segments retained
from the data. As we can see, all p-values are rather low, indicating that, overall,
each of the six clustering variables differs significantly across the clusters. Note that
this does not mean that the clustering variables differ between all segments – this
result merely indicates that at least one cluster is significantly different from the
others with respect to each clustering variable. In order to evaluate whether all three
clusters exhibit significant differences, we would have to carry out pairwise compar-
isons using post hoc tests (compare Chapter 6). Even though this analysis renders
interesting results, we are primarily interested in the F-values (second column from
the right) which partly correspond to the VRC statistic (Compare Chapter 9).
Table A9.2–A9.5 show the ANOVA outputs for a three- through six-segment
solution using k-means.
Table A9.1 ANOVA output for k-means analysis with two segments
ANOVA
Cluster Error F Sig.
Mean Square df Mean Square df
moment 122331,868 1 1208,902 13 101,193 ,000
width 47761,376 1 1818,643 13 26,262 ,000
weight 668960,744 1 10833,584 13 61,749 ,000
trunk 212487,619 1 9598,593 13 22,137 ,000
speed 14994,525 1 415,990 13 36,045 ,000
acceleration 83,160 1 3,326 13 25,004 ,000
The F tests should only be used for descriptive purposes because the clusters have been chosen to
maximize the differences between cases in different clusters. The observed significance levels are
not corrected for this and cannot therefore be interpreted as tests of the hypothesis that the cluster
means are equal.
4 Application of Variance Ratio Criterion (VRC) by Calinski and Harabasz (1974)

Table A9.2 ANOVA output for k-means analysis with three segments
ANOVA
Cluster Error F Sig.
Mean Square df Mean Square df
moment 62902,371 2 1020,238 12 61,655 ,000
width 23885,010 2 1969,476 12 12,128 ,001
weight 337371,095 2 11254,595 12 29,976 ,000
trunk 142820,381 2 4302,381 12 33,196 ,000
speed 7655,200 2 424,333 12 18,041 ,000
acceleration 44,975 2 3,037 12 14,808 ,001
The F tests should only be used for descriptive purposes because the clusters have been chosen to
maximize the differences between cases in different clusters. The observed significance levels are
not corrected for this and cannot therefore be interpreted as tests of the hypothesis that the cluster
means are equal.

Table A9.3 ANOVA output for k-means analysis with four segments
ANOVA
Cluster Error F Sig.
Mean Square df Mean Square df
moment 34064,533 3 3259,455 11 10,451 ,002
width 16058,400 3 2111,685 11 7,605 ,005
weight 244397,944 3 6963,955 11 35,095 ,000
trunk 97282,878 3 4129,155 11 23,560 ,000
speed 3854,400 3 803,564 11 4,797 ,023
acceleration 23,753 3 5,012 11 4,739 ,023
The F tests should only be used for descriptive purposes because the clusters have been chosen to
maximize the differences between cases in different clusters. The observed significance levels are
not corrected for this and cannot therefore be interpreted as tests of the hypothesis that the cluster
means are equal.

Table A9.4 ANOVA output for k-means analysis with five segments
ANOVA
Cluster Error F Sig.
Mean Square df Mean Square df
moment 25989,233 4 3409,067 10 7,624 ,004
width 12642,333 4 2083,440 10 6,068 ,010
weight 188779,467 4 5467,947 10 34,525 ,000
trunk 73369,167 4 4379,267 10 16,754 ,000
speed 2974,133 4 850,587 10 3,497 ,049
acceleration 17,917 4 5,473 10 3,274 ,058
The F tests should only be used for descriptive purposes because the clusters have been chosen to
maximize the differences between cases in different clusters. The observed significance levels are
not corrected for this and cannot therefore be interpreted as tests of the hypothesis that the cluster
means are equal.
References 5

Table A9.5 ANOVA output for k-means analysis with six segments
ANOVA
Cluster Error F Sig.
Mean Square df Mean Square df
moment 26402,460 5 670,589 9 39,372 ,000
width 12188,247 5 1162,500 9 10,485 ,002
weight 156543,607 5 3008,811 9 52,028 ,000
trunk 65755,047 5 943,789 9 69,671 ,000
speed 3675,120 5 225,200 9 16,319 ,000
acceleration 22,554 5 1,514 9 14,895 ,000
The F tests should only be used for descriptive purposes because the clusters have been chosen to
maximize the differences between cases in different clusters. The observed significance levels are
not corrected for this and cannot therefore be interpreted as tests of the hypothesis that the cluster
means are equal.

To compute the VRC statistic for each case of number of segments, we simply
sum up the F-values from Tables A9.1-A9.5, using a spreadsheet program such as
Microsoft Excel. The results are shown in Table A9.6

Table A9.6 VRC values Number of clusters VRC

2 272.390
3 169.804
4 86.247
5 71.742
6 202.770

To determine the “correct” number of segments, we compute oK for each

segment solution. For example, for K ¼ 3, oK is given by

o3 ¼ ð86:247  169:804Þ  ð169:804  272:390Þ ¼ 19:029

Similarly, we can compute oK for four and five segments resulting in o4 ¼

69.052 and o5 ¼ 145.533, respectively.
Comparing the values for oK, we establish that the minimum is achieved for
K ¼ 3. Thus, we would choose a three-segment solution for our analysis.

References

Calinski, T. and J. Harabasz (1974). “A Dendrite Method for Cluster Analysis,” Communications
in Statistics – Theory and Methods, 3 (1), 1-27.
Milligan, Glenn W. and Martha Cooper (1985). “An Examination of Procedures for Determining
the Number of clusters in a Data Set,” Psychometrika, 50 (2), 159-179.