Graphics Reference
In-Depth Information
a large bar means a large number of votes. An alternative would be to use differences
between cluster centroids and population centroids instead of absolute values; see
also Fig.
.
. For example, cluster one has only slightly above-average results for
all parties, but the LINKE result in cluster three displays a large difference from the
total LINKE mean. Note that one cannot easily test whether the mean of a cluster
is significantly different from the mean value of another cluster (or the population
mean), because the clusters are not independent of each other and were constructed
to be as different as possible from each other. If cluster dispersion is also of interest,
we could replace the bars in Fig.
.
by boxplots of the points in each cluster.
Parallel coordinate plots (see Chapt. III.
, Inselberg) show all variables of a high-
dimensionaldatasetinonefigure.Clustermembershipcanbemarkedusingdifferent
line types and colors for the clusters. his will work for data sets with relatively few
observations and/or well-separated clusters with small within-cluster variation. Fig-
ure
.
shows a trellis display (Becker et al.,
) of parallel coordinate plots where
each panel corresponds to one cluster. his approach works well for a large number
ofdata pointsandclusters,because clustersarenotplottedovereachother.Ifallclus-
tersareplottedonthesameplot,linesdrawnlaterbythesotwarepartially masklines
drawn earlier.
Figure
.
.
Parallel coordinate plot of the five clusters
