Page 141 - Jolliffe I. Principal Component Analysis

5. Graphical Representation of Data Using Principal Components
110
its largest coefficients are all positive and correspond to numbers of elderly
persons who have recently moved to the area, numbers in privately rented
accommodation, and population sparsity (Area/Population). The implica-
tion of the outlying curves for Cluster 12 is that the two local authorities
corresponding to the curves (Cumbria, Northumberland) have substantially
larger values for the seventh PC than do the other local authorities in the
same cluster (Cornwall, Gloucestershire, Kent, Lincolnshire, Norfolk, North
Yorkshire, Shropshire, Somerset and Suffolk). This is, indeed, the case and
it further implies atypical values for Northumberland and Cumbria, com-
pared to the remainder of the cluster, for the three variables having the
largest coefficients for the seventh PC.
Another example of using Andrews’ curves to examine the homogeneity
of clusters in cluster analysis and to investigate potential outliers is given
by Jolliffe et al. (1980) for the data set discussed in Section 4.2.